-
Notifications
You must be signed in to change notification settings - Fork 0
The Reference (whole)
-
AI Dungeon World Info Research and Reference
- Preface
- Recommendations or the TLDR
- Remember WorldInfo AuthorsNotes Worlds and how the game handles them
- Griffin vs Dragon
- Tokenization - understanding limitations and special characters
- On certain characters
- Authors Notes specifics
- World Info and Formatting
- Other Scenarios and Scripts
- Writing Tools
- Obsoleted Methods
- Special Inputs Commands
- Ending Notes
This document contains a lot of examples in code blocks, with very long lines of text. If you're reading this from GithHub website (I bet you are), you might want to add this CSS style in your browser to enable word wraps:
code {
white-space : pre-wrap !important;
}You can also use Dark Reader or similar browser extensions to switch GitHub to a dark theme.
AI Dungeon became a truly great storytelling assistant after the addition of the world info feature. Yet even after more than half a year of using it, many people are still unaware to its power, potential and effective use. Learning and figuring out how to use world info effectively is of great interest to many users: using it one can get vastly more consistent output from the AI and define terms that the pre-trained AI might otherwise have difficulty understanding.
Learning about WI from scratch can also feel like an overwhelming task, something this document seeks to help with. While the document welcomes fesh readers, it's written with current and future enthusiasts in mind. Throughout the document we take a thorough look at the things going on below AID's surface, explaining why and how things work the way they do, giving the reader a comprehensive overview to build personal world infos on.
This document flows a certain way due to the history of work the enthusiasts put into developing their methods. We respect the reader's intelligence and as such no effort will be made to dumb down the subject matter. Every segment is important for understanding how to write for AID. Effort has been made to present information concisely and in a straightforward manner. Accept no cut-down alternatives.
Features like Remember (pins) and A/N: (author's notes) will also be covered as they are heavily related and connected to world info. Fundamentally each tool Latitude has given us is about manipulating the context of the game. Everything here is of use to both free and premium users. The differences between griffin and dragon will be discussed as there are some differences one should be mindful of.
Something that we need to be clear about: This document isn't authoritative and doesn't pretend to be so in any section. It's a compilation of the community's shared and useful experience, primarily through the official AID Discord and burrowing publicly available material from other fan-run communities. Originally it was the author's personal notes and should be treated as such.
The above means the evidence and advice is mostly anecdotal. Citations are shared with permission from contributors. Most of them can be found as messages on Discord. Where relevant examples are provided as screencaptures. Everything has been repeatedly tested inside AID scenarios and the tools provided to us by developers and scripters. Some outputs will be skipped for the sake of brevity as everythig can be replicated and tested inside AID. Do your own research and testing. If you fail to do so, criticism won't be taken seriously.
As AID evolves so does this document alongside it. One of the changes included and discussed is scripting that is now free to all users. Written documents from other authors and their work will be provided with permission. For future-proofing, we've included a category briefly going over past methods in case they may one day be useful again.
Special credit to kimtaengsshi for help with editorship and fact checking the technical details. This document will contain data and public statements from kim and other trustworthy individuals such as one of the developers, Matt. Much of this information comes from and is shared with the community-run wiki and Discord. A link to the wiki will be provided where you can also find an invite to the Discord and other useful content: https://wiki.aidiscord.cc/wiki/Main_Page
If you're a fresh AID player don't know the first thing about world info or other advanced features yet it is recommended you read the wiki article first. AID even includes a tutorial inside the game itself now: https://wiki.aidiscord.cc/wiki/World_Info
This document isn't anti-prose, but prose itself has little to do with the subject matter and as such won't be focused on. This document is about getting results by manipulating AID's context using any method available. We have a problem-solving approach to issues discovered in modifying AID's context and output. There exists a countless amount of other resources online to help you become a better writer overall which should also improve your AID experience if you seek them out.
The owner of the repo(valahraban) is pseudonymously referred to as the author or simply birb due to personal preference. How to get featured or help improve the research sheet? Send a text file (preferably .md) or suggested edits in a pastebin variant to birb/valahraban through github or discord. This is how he primarily communicates with other enthusiasts and writes this document. Other communication methods aren't supported. Included information isn't chosen at random: it must be novel and immediately useful to birb and WI enthusiasts. Some aspects of writing have been left out of the document precisely because they lack useful utility in the context of AID and WI.
Work on this project is an endless work-in-progress and moves on at the author's own pace. Recent events have shook this pace. If you desire more up to date new tricks, it is recommended to visit other guides like ones by Rinter that have also been linked within this document.
Scripting has been made free for all users, it is warmly recommended you scroll down to Zynj's github and check out EWIJSON. Loading WI through these scripts alone gives you a lot more control over their location. If you do decide to use EWIJSON, also check out Monky's Futureman guide and birb's guide on EWIJSON basics inside the docs. This isn't mandatory. This document will focus on scripting-free methods. Scripting focused documents are included in the docs folder and mentioned where relevant.
This is for those of you who came here for the WI formats or have short attention spans. Click on the links contained in the table of contents above or hit CTRL+F to search for names that interest you.
JSON formatting is neither recommended or discouraged. If you're comfortable using it continue doing so. The Zaltys Formatting of this document is recommended as a strong baseline to start from. Zaltys formatting is easy, popularized and frequently updated to keep up with AID. Past samples are provided as well. Monky Formatting is recommended for people with specific goals and interests such as intending to publish scenarios, worried about WI space and token use and wanting to share their work. It is very understandable, easy to modify and works well with regular prose. Very easy to combine with scripts too with included documentation. Onyx has written formats and WI tricks for specific purposes such as defining locations and listing things. Her original examples are somewhat dated and birb has tried updating them.
We now use the term formatting as some of the concepts are interchangeable and may be combined together. One scenario may also use different WI formatting for achieving different things with different entries. You the writer have complete creative freedom, although our experiences may benefit you.
If you don't care about the complicated lunacy of this document and want something simpler, more elegant and less of a headache, learn how to write good prose or read a simple guide on it! Now provided by our fellows at /aidg/
https://rentry.co/aidg-format
Check out Rinter's clearly written wiki for other tips and discoveries on how to best structure WI. It serves as excellent supplementary material to this document especially if you want to get into using scripts.
https://github.com/RinterWaptor/AID-research/wiki
For comedic purposes here is what one very opinionated user has to say on the current state of formats:
From CrisAIcilian: Zaltys is better in every way, it's actually designed for its purpose. JSON is like Chinese medicine, goo of mumbo jumbo happens to contain three elements that had an effect (colon, comma, square brackets).
Remember was the first of these features implemented in the game. This segment goes into technical details so you may want to skip it if you're here for formatting World Info. This section is very important if you want to understand AID as a tool and it's limitations. It will also answer most problems people have with their WI not working. Grossly simplified remember is a world info that gets constantly fed to the game after your every input. World info is like remember, but has user-defined key-words keys that after getting triggered make the game use them as part of the context. Keys use string-matching which will be important later. Having this element of context makes world info less restricting to work with because when you're happy with it you can forget and leave it while things in remember require more frequent attention and updates as the story progresses. History means old prompts written by you and the AI. The order the AI reads your input is in:
- Summary | Experimental/deprecated feature to help devs
- WI
- Remember
- History
- AN if you've passed exactly 3 newlines in history
- Rest of history
- Fresh Input
- Front memory from scripting
The following data is based on information available to scripters or based on their limit testing, documented on discord. All of the above consumes characters or as we will soon discuss tokens from the AI's memory. Anything that goes over the memory limitations will not be read by GPT-3 and is therefore irrelevant to the output process. The entirety of the data(text) that gets fed to GPT-3 is called the context. The following is a list of rules the context follows as related to the many parts that form it:
- Max. input character length: 2772
- Max. Summary length: 500
- If (Summary + WI + Remember) < 50%, then extra story aka history will fill up the empty space
- If (Story + A/N + Input + frontMemory) < 50%, the unused space will not be allocated to (Summary + WI + Remember)
- GPT-3 has a hard cap of 1024 tokens for AID users, 2048 for the devs
For brevity the maximum context length is often shortened as 2.8k. In regular use a single WI has a limit of 500 characters and remember has a limit of 1000 ch. On the plus side thus far no limits have been discovered on how many WI one can define. It's not recommended, but it is possible to define multiple WI with the same key, essentially increasing the space by 500 ch on every iteration. Keep in mind AID always adds a newline between each WI even those using the same key.
AID is smart about using WI, but not about remember and updating remember constantly is tiresome therefore many advanced users focus on producing high quality WI while forgoing longer remember pins. The limitations of GPT-3 and AID is why understanding and researching WI is of interest. Worlds will be discussed later as they use the same data structures as regular AID. Going by the formula above, 3 WI is cited as a common limit, because history is usually taking up about 50% of the context, leaving you with half(1336) characters for WI and Remember.
With scripting it's possible to circumvent the regular ordering of each element. This is a topic that the user should keep in mind if they are interested in scripting and the potential benefits.
AID contains multiple machine learning models the primary ones being Griffin and Dragon with Lovecraft being a later seasonal addition(now phased out). Modern versions of AID Griffin and Dragon are both GPT-3. This has been confirmed by the devs on twitter, discord and other public forums such as their medium. That being said there are different sizes of GPT-3. Griffin uses a subset of GPT-3 that is closer to the original 1.542 billion parameters used by the GPT-2. The full-sized GPT-3 that Dragon is based on has a whopping 175 billion parameters. This shows in the quality, consistency and most importantly forgiveness of the model and the output it produces. Both models benefit from a detailed and consistent context, but Dragon is more forgiving in terms of formatting. This shows especially in how differently it responds to WI you feed it. Griffin has quirks you have to be mindful of. These affect what kind of WI you write and will be discussed in the following sections.
It used to be that the first prompt and response were generated in GPT-2 but this is no longer the case, it depends on what model you use. If there is a populated prompt, the first output will be in whatever model you're using. Only a completely empty prompt would be GPT-2. So at minimum put something in your prompts. This is official information from Alan. Both detailed history and WI are important for training the AI for your scenario and getting what you want out of it. Either redo until the writing looks high quality enough to your liking or spice it up with attentive alters, the key to enjoying AID is building on success.
The character limitations mentioned in the previous section are ones imposed by the devs of AID. They are quite sensible, plenty for generating good stories, not bankrupting Latitude and not horribly far off from the true limitations of GPT-3 that both models now run on. For a machine learning transformer to read an input it must first be tokenized into a form the model understands. GPT-3 has a hard limit of 2048 tokens. It won't read more at once. Devs have access to 2048 tokens, the users are given 1024 per input. But as you've noticed, you have more words to work with that the AI clearly understands. Tokens are just the way data is encoded into the machine learning model; during training, common words became recognized and encoded into an efficient form, sometimes even a large word gets recognized as 1 token. From kim's research, depending on how a world info is defined, the counts average to 2-4 characters per token meaning you get to use at least double the characters of the tokens you really have hence why a commonly mentioned character limit is 2048.
A few things have been learned from experience; common words and phrases like hello or 25y get treated as a single token. White spaces are special because they often get rolled into words as is grammatically correct. An uncommon word like Enterprise takes 1 token with the white space while Enterprise takes 2 tokens. Unicode emoji like 😃 and some special symbols are expensive, they cost 2 tokens per character. As we will discuss characters like <> [] are useful for WI, but tend to have a cost ratio of 1 token. There are also rare characters that the AI hates like ù. Experiments have shown that longer strings of symbols like }}} get rolled into 1 token.
If you want to test out and learn more about GPT-2/3 tokens (they work the same) visit this link: https://colab.research.google.com/drive/1i1wXfc4SWDprHZuLTaMsgWPv6YwvlQ_I?usp=sharing#scrollTo=mlBq9zO67b2n
kim has also made a tokenizer scenario/script you can use inside AID: https://play.aidungeon.io/main/scenarioView?publicId=6927d610-34b7-11eb-b8e1-c185d026672d
Tokenstuff script by Gnurro that can be used to mass search and catch what tokens exist inside GPT-2. Remember, by the word of the developers(God) GPT-2 and GPT-3 share the same tokens: https://github.com/Gnurro/AIDscripts/blob/main/tokenStuff.zip
It can not be understated what an important tool the tokenizer has been for WI enthusiasts. Thanks to it we have a fairly good idea why certain words work as poorly as they do and a legitimate information-based argument for why this is the case. If the document talks about word or character tokens, they were ran through the tokenizer colab or scenario alongside synonyms or longer compound words containing it.
A common misconception is that """natural writing""" is clearly the best writing method and all other methods and tricks should be ignored because some marketing person said so. To these people I must tell: Stop. You don't know what you're talking about and are likely arguing from emotion. Machine Learning language transformation algorithms expose their training biases through tokens. You can check the tokens through the tokenizers. They are the same for GPT-2/3 and the Griffin and Dragon model trained off GPT-3. The GPT-3 has eaten a massive amount of online datasets, not restricted to plain English. Tokens exist for html comment wrappers like <!-- and -->. This means they were a relevant part of the GPT-3 training and will be recognized by any AI program using the same algorithms. We now have a free web service you can use to rapidly test how strings tokenize for GPT-3.
https://ermela.net/gpt-3-encoder/
Input the string SolidGoldMagikarp into the tokenizers above. This nonsense word is one token I kid you not. One possible explanation for this we've found is reddit.com/r/SolidGoldMagikarp/ although the true source of the token remains unknown. ================================ and many other chapter breakers are one token too. This means GPT-3 strongly recognizes these strings and gave them a unique definition inside of its tokens. The AI has eaten data from a plethora of writing sources. You can further train GPT-3 to specialize on other forms of writing, as Latitude did with CYOAs for the AID models. That is an official statement. The author would have preferred they focused on high quality prose novels like the Witcher series or Dune(copyright problems), but we live with the tool given to us. We provide pictures of two posts by Gnurro as he researched the topic of unusual tokens:
https://files.catbox.moe/xj9a6a.PNG https://files.catbox.moe/qz4hwx.PNG
Any section of this document discussing tokenization has been verified with either of the above tools. They are the basis for why some tricks used by enthusiasts and their formatting work.
Functionalities of certain characters have been found and described thanks to the development of world info formatting but it's useful to discuss these beforehand. Their functionality can be somewhat verified through the tokenizer Colab. Before we proceed, keep in mind the document was started in September of 2020. Updates to OpenAI's back-end can and have changed the meanings of characters in the past.
Some characters connect. The most important of these used to be -. hair-color or red-color are words that got better results for colorization than other methods before January 2021. After this update, many characters including - started being treated more like math symbols by the AI. Often you can even string random words together which will be discussed later. Enthusiasts have also found that _ is good for creating a strong connection between words, especially for categories. Some are also experimenting with . for connecting categories. Onyx has provided us with examples of the AI's math storytelling ability.
https://files.catbox.moe/8y17ji.png
The functions of special characters are mostly dependant on what OpenAI does with the GPT-3 backend. As of April 2021 it's safe to use - in WI entries again, especially if the word is encapsulated. This indicates that OpenAI has continued fine-tuning to make characters like - more context dependant. The best advice we can give the reader is to test out their entries first, especially if guideds haven't seen updates for over a month.
AID recognizes many separators. Letters , '' "" ; / | all separate words and concepts in rough degree of power. You would want to separate members of a list with , like in common grammar, but ; might be better for separating categories or ending semantic structures. Inside brackets / is useful for keeping the traits related to what they are describing. () {} [] <> : are connectors. You can put lists inside all of them or use : to create a logical implication. For example it might be preferred to do a list like MENTAL:[happy, go-lucky, energetic]; and define logical connections like AGE(25y) or sword(polished) the AI usually recognizes that the thing you're describing is roughly 25 years old or that the sword is polished. Enclosures can be interchanged, but these are two commonly seen and powerful methods. For list specifically, & is powerful to get the AI to mention everything connected with & together. It is likely, but untested that | works more like logical OR now too.
In the January 2021 update we mentoned before () became worse at encapsulation, but may be used for logic connections. As of April 2021 OpenAI seems to have further fine-tuned the use of (). It isn't as useful in entries as it was in the past, but inside the story/context of AI parentheses are useful for adding extra information that isn't necessarily apparent to all the characters in the scene. Fair warning, if combined with front-loaded scripting, using a lot of parentheses encourages the AI to keep frequently outputting them. This can be useful or detrimental depending on your writing style.
< and > are special characters. You may have noticed in do or say mode the AI always prefaces your inputs in the story with >. In AID the character is reserved for this purpose. You may have noticed that the AI never outputs < or > on its own. Since the information inside WI is handled differently we can still use any unicode letter inside it. In WI this can be used to our benefit as we will see when we discuss the work Zaltys has done and his formatting. The uses we have discovered for < and > are particularly useful, because they are unlikely to change even with future updates unless Latitude decides to drastically change their AI model.
All quirks of the AI are due to how GPT-3 was trained as much of the above illustrates. These may also change as it is a work-in-progress type of application. Some particularities are introduced by Latitude's training and weighting parameters. One often acknowledged example of this is the appearance of Count Grey and other characters from the Blooded CYOA. The following article is a good read on understanding this topic.
https://aidungeon.medium.com/controlling-gpt-3-with-logit-bias-55866d593292
Let's discuss the premium feature A/N: before the other features. If you're looking to inject flavor into the AI's outputs A/N will be your first and probably most efficient stop. It is a worthwhile for the convenience, yet its usefulness is not entirely restricted to premium users with our advice. What A/N does is it insert a line of text into the history the AI reads e.g. Author's Note: Some details like mood about this story. exactly 3 newlines up. On the scenario screen it has a 150 character limit, but the in-built maximum limit is 300 characters according to kim's LMI testing and community experiences.
Discovering manual Author's Note inputs during dev testing and how heavily it affected output is why it was added to the core game. Even as a free user you can manually add/edit Author's Notes into your story. With manual A/Ns you have more characters to work with, but it is doubtful that making a super long A/N: offers any extra benefit.
Research suggests that A/N: can be used to reinforce writing style even of well-known authors through repetition. It can also be used to change perspectives and the focus of the storytelling. These findings suggest that the prompt informs the AI what kind of writing to expect directly after A/N: has been invoked. Exemplified below is a medium-length A/N: with reinforcing that has been confirmed to be capable of producing outputs where characters break the 4th wall:
[Author's note: This novel is esoteric and descriptive. Author: Terry Pratchett. Writing style: Metafiction, in the style of Deadpool. Genre: Witty, talkative.]
The film rating system (not others) has been confirmed working. With RATING: G in A/N you can noticeably lower the amount of NSFW in your scenarios. Alternatively RATING: R makes things more racy. X-RATED works too with predictable results. As with most things this is more effective on Dragon. Arguably an even more effective method of reducing NSFW is using Author: and a known author who considered profanity to be rude such as Samuel Clemens or Edgar Allan Poe.
Switching focus is easy, but may require extra use of alter afterwards to make sure the AI understands what you want. This style of A/N: is best used manually written into the story (although you can change your global A/N whenever you want; a trick we recommend using). [Author's note: The story now follows Mary's perspective.] or [Author's note: We now switch focus on Mary.]
The main character can also be defined in A/N vs memory if that is what you desire A/N: Main character: <species>, <gender>, <job>
A/N likes both style-describing nouns and adjectives. strikingly elegant makes the AI output more elegant prose than just elegant. This may also work for other words predicated with strikingly. THEME: is a word that can get the AI to focus on concepts and their attributes such as pirates or ninjas.
An example of output generated with strikingly elegant style:
You look around.
The sun softly kisses your exposed skin with her rays. You gently bask in her glory, raising your face towards her. Below, the plants softly sway in a gentle breeze as though worshipping her as well.
List of common categories to throw inside Author's Notes: author, writing style, genre, theme, setting, scene, format, goal, situation, storyline.
A/N understands JSON too. Its entirely possible to do something like A/N: [{"writing style":["descriptive", "elegant", "gritty"], "wording":["archaic", "Cockney accent"], "current state":"indoors"}]
Some users have been experimenting with A/N for doctoring output and specifically managing randomness. Lower randomness is more consistent and predictable as it makes the AI's predictions more deterministic, but this comes at the expense of it feeling 'more dumb', hence higher randomness being desirable for verbosity. According to CrisAIcilian This novel has no plot twists, it follows a linear storyline. is an effective phrase to include in A/N to increase consistency. This way his stories can stay on track even at 1.3 randomness. He finds that Writing style: and Genre: are useful categories to use most of the time. Writing style improves writing, genre works like a mini-prompt. More useful notes on writing A/N from Cris:
"Author:" is added when you want to do a convincing real author (combined with "Writing style: literary, author-name.) "Setting:" "Theme:" "Subject:" have their uses, depends on the content. "Title:" works at least on LOTR and "gone with the wind". Didn't test other titles. Famous ones should work. Has to be a real title, "star wars" is not a title, it's a movie franchise. The writing styles, most don't work on Griffin.
Only the descriptive-branch works on Griffin. "descriptive" works, "narrative" and another one I can't remember (according to Zaltys) also works on Griffin.
After months of playing around with Author's Notes the world-info discord community mostly agrees it works like any other bit of WI entries, but inserted exactly 3 lines up. We've discovered that all categories using the word notes tends to be pretty powerful and a category that works very well to alter dialogue Dialog notes:.
While less effective on Griffin anyone can benefit from using Author's Notes properly. With scripting, any user can mimic the functionality of Author's or Editor's Notes or even insert any string they want in a specific part of the context. Zaltys is experimenting with Author's Notes words all the time and has provided a resource that's being updated frequently: https://justpaste.it/9ofj1
The section most users will be interested in. For the longest time users of AID were writing WI in pure prose. It helped things, but things remained inconsistent especially with how fickle Griffin tends to be. Users started researching using different letters and sentence structures. Some people even tried coding inside WI or getting AID to code for them. After some time, users shared their results and many came to the shared conclusion that deliberate world info formatting leads to more desirable outcomes. These became known as world info formats. Some goals of structuring WI include:
- Reducing randomness (or later adapted into a RND generator)
- Defining characters, locations, lore AND getting the AI to remember and mention these things
- Describing those things consistently and repeatedly
- Staying on topic: more magic and less guns in fantasy worlds, no magic and more scientific guns in sci-fi worlds
- Preventing leakage - the AI picks up on patterns and WI can leak (especially in Griffin)
- Consistency both in how the AI receives data and how the user writes their WI (the AI learns from consistency, consistent WI are easier to write)
- Formats can be mixed and matched for different tasks; characters, new verbs/actions, locales and regions might each have an optimal format that better suits the user's needs
- Breakdown: sloppy WI and sloppy writing can lead to the AI 'losing focus' faster and forgetting about what you're trying to tell it with WI (occurs faster in regular stories)
One of the first formats commonly shared online was popularized on Reddit by a user curious_nekomimi in August 2020. The format had no common name, but each new line of the example started with Ben so that's what people usually called it. Every newline would start with the key being mentioned and every sentence would be followed by a newline. This raised public awareness of reinforcing the AI with WI. A simple prose format like this can get the AI to be mindful of most attributes assigned to said character or object. But it tends to be heavy on used-up characters and especially Griffin might start mixing up actors from different WI together, now called leaking.
Before we go further let's discuss a common, yet counterintuitive pitfall of WI use. That is defining the main character or you in keys and ultimately the WI entry. This makes sense in short scenarios focused on the main character and a small group of characters that have one-on-one interactions. But if we use many naive WI keys in a longer fantasy adventure with many unique characters, we notice the AI paying less attention to the information you've given to each entity inside WI.
This is due to WI overloading going by the rules we discussed at the beginning of the document. When simple keys like you are used they are active almost all the time, taking up precious space which is restricted to roughly 1386 characters going by the rules. If we use a fully stacked A/N and WI entry for you, we've used up 650 of the possible 2772 maximum. When you have WIs for the location and multiple characters, these space limits get eaten up fast, leading the AI to ignore some of your WIs. Enthusiasts generally avoid keys that are active too often for this reason. They are best used as a strong reminder when needed, the shorter the better.
It is often preferred to define the character you towards the bottom of the Remember pin as that gives you the fullest amount of control where those details get inserted. No best practice exists for this type of WI, so it is left to the reader to deliberate on what their needs for AID are.
The author advises against flooding the world with more format names. This serves mostly to only confuse new players. Originally formats were a quick way to reference different directions of WI research. Now the different methods are starting to converge back as the enthusiasts are attempting to optimize their personal stacks and styles. Formatting is a much more accurate term and beneficial to everyone in the community without excluding any form of manipulating language. This document hopes to demonstrate this.
Fun fact: format: works extremely well inside Author's Notes as a keyword/category you can use to tell the AI whether you're writing a script or a sonnet or a manifesto, etc.
WI can be exported and imported out of AID as JSON files. What we discuss here is something different. Interest in WI formats blew up after people started publicly sharing anecdotes of a JSON-like WI format having great results in August 2020. Many people did private research at the time but Zaltys 🐍#5362 is often credited with raising public interest. Here we define the pure JSON format:
keys: AKA: you,Anon
Entry: [{"You":{"name":"Anon","age":24,"gender":"male","species":["human","man"],"eyes":{"eye-color":"brown"},"hair":{"hair-color":"dark brown","hair-length":"intermediate","hair-style":"spiky"},"relationships":{"orphan":"no relatives"},"body":{"physique":"average","height":"average-178cm","weight":"average-80kg"},"personality":["sarcastic","friendly","wishes he had a gf","plays AI dungeon"]}}]
Spoiler: Expanded version
The same entry, with tab indents (don't put it into WI entry, it's here just for readability):
[{
"You": {
"name": "Anon",
"age": 24,
"gender": "male",
"species": [
"human",
"man"
],
"eyes": {
"eye-color": "brown"
},
"hair": {
"hair-color": "dark brown",
"hair-length": "intermediate",
"hair-style": "spiky"
},
"relationships": {
"orphan": "no relatives"
},
"body": {
"physique": "average",
"height": "average-178cm",
"weight": "average-80kg"
},
"personality": [
"sarcastic",
"friendly",
"wishes he had a gf",
"plays AI dungeon"
]
}
}]
The many things we have learned about using JSON inside WI can be summarized with the following list:
- Super consistent, the AI understands it well and tends to stick on point with it
- Simple syntax, making it easier to produce for non-native English speakers
- No reports of leakage (on discord)
- Worked even during the AI-issues that happened during the cache-bug fixes(called the Great Stupid in the community)
- If you know how to write JSON you can describe more complex relationships using the syntax and have the AI use them
- Consumes many tokens, space/memory wasted on "" and other symbols
- Motivated research into other formats because of it's usefulness and caveats
- Painful to write and spell-check, use a linter/editor like jsonmate.com and validate with jsonlint.com or use a coding text editor
- You're best off writing minified JSON because styled/linted JSON has no benefits, uses more memory space and looks weird when you paste it into WI or inside the proper JSON-file
- In Discord community testing reduced JSON without quotation marks produced mostly similar quality output
JSON had its uses. If you're already using it feel free to continue doing so. If it looks frightening, don't feel obliged to invest learning it. Most WI users have moved on to using other methods specialized for their purposes.
Zaltys 🐍#5362 has popularized many interesting formatting methods and spurred the community to try many different things out with the tools AID provides us. Because originally there were few enthusiasts trying unique tricks with the world entry screen, Zaltys distinct style of formatting text became simply known as Zaltys and was adopted so people immediately knew what was being discussed. This document will be kept up to date with the most recommended format variants of Zaltys with some explanation of his goals and motivations. First we have standard Zaltys that works with any model.
Mike Haggar:[Human<male, 202cm, 140kg>. APPEAR<Haggar>:Stocky, muscular, big arms, hair<brown>, mustache<brown>; MIND<Haggar>: Just, upright, direct, upbeat; WORN:Eyeglasses, pants:<green>; SUMMARY<Haggar>:Born in 1943, from Final Fight and Street Fighter games, ex prowrestler, grew up on streets in Metro City, mayor of Metro City, fought Mad Gear & Skull Cross gangs, fights gangs, "It's my job to keep Metro City safe!", loves curry, friends:<Cody, Guy>, daughter:<Jessica>.]
- This format must have a . near the end to work correctly. Do not use extra periods within categories, not even in quotes.
- Avoid using -s. (Jan 2021 change.)
- Including the name (inside
<>) after the categories helps Griffin parse the entry. - Categories must be typed in UPPER CAPS, with the exact spelling.
APPEandSUMMcan be used as abbreviations, most other abbreviations do not work properly. - Other useful categories include
MOV(for movement-types, such as 'slithering'),GRAB(for object manipulation, such as 'talons', 'beak', or 'coiling'). - If you need to add age, use age<num> in same section as weight/height. Ages get more unreliable after 20s, I recommend rounding them to nearest five-dividable age (30, 35, 40, etc.)
- All of it goes into single WI entry with no line breaks, under a keyword such as
haggar.
We also have an alternate advanced version specifically designed for Griffin. It saves space inside the WI and context, but is both harder to read and write. Advanced alt format ('Snek', upd. Feb 2021):
Mike Haggar:[Human<male/202cm/140kg>;APPE<Haggar>:Stocky&muscular/bigarms/brown<shorthair&mustache>;MENT<Haggar>:Just/direct/upbeat/hardy/upright;WORN:Eyeglasses<at work>/brownboots&greenpants/bandolier;TRA<Haggar>:Born 1943/from:<Final Fight&Street Fighter> games/ex:prowrestler/grew up on streets<in Metro City>/mayor<Metro City>/fought gangs:<Mad Gear&Skull Cross>/fights gangs/"It's my job to keep Metro City safe!"/loves:curry&rice/friends:<Cody&Guy&Carlos:from Brazil>/daughter:<Jessica>.]
- Saves space, better accuracy on Griffin.
- Harder to write and understand, confusing usage of spaces.
Zaltys combines most WI writing tricks the community has discovered into one very efficient package. One caveat would be how strictly you have to follow the above patterns to guarantee functionality.
Based on experiments and accumulating experience the community has found all-caps categories are useful. Later on in the document we discuss categories in depth and include a long list of them that have been discovered to be useful for any WI format desiring to utilize them.
Zaltys likes to write a lot of species from different series. He has written on discord that his most used categories are APPEAR, LACK, GRAB, MOV, MENTAL and TRAITS later superceded by MIND and SUMMARY. For defining a type of character all these have purpose. APPEAR is self explanatory. LACK has to do what the characters can and can't do; a lamia without legs shouldn't be able to walk. GRAB is their preferred way of interacting through appendages, a bird would use it's beak or talons, a lamia would use their hands or tail. MOV works for movement. Lamia would crawl/slither on top of their tails, slimes ooze and other species may do more imaginative things. MIND is traits of the mind, SUMMARY is more intrinsic physical, sensory or nature related traits or important things of the character in general.
TRAITS proved to be the Swiss army knife of categories after months of use, but later SUMMARY proved to be even better for generic use. INV and EQUIP work similarly to one another as alternatives to WORN and can be used to define the types of equipment a RPG character would carry on top of the usual stuff like clothes. WORN is better at handling style like goth clothing. gothic is more applicable to architecture.
How Zaltys defines you inside the remember pin:
You are male human named Zaltys.
[ you.DESC: adult, gamer;
you.INV: wallet, keys<in pocket>;
you.MIND: confused;
you.LOCATION: floating in outer space.]
The AI is good at overwriting preset associations. Following Zaltys examples the community has succeeded in writing interesting recipes. For some reason the AI seems very good at describing recipes. This is another example of an indicator helping the AI understand WI better. Example from Mono:
Siren:[A popular drink;DESCRIPT<Siren>:Sapphire-colored/silvery bubbles/sprig of mint/smells alcoholic/tastes like rusty metal;EFFECTS<Siren>:Inspires drinker to dance.]
> You describe the siren to Severn.
"It's a drink," you say, "a sapphire-colored liquid with silvery bubbles that smells of alcohol and tastes like rusty metal."
"Sounds poisonous," says Severn.
"No. Not at all. It's delicious."
"You've had it before?" asks Severn.
You nod.
> You tell him about the side effects.
"It inspires people to dance," you say, "my sides hurt after I've had it."
"What else?" asks Severn.
"That's it," you say.
"That's all of it?" asks Severn, "no other effects?"
You shake your head.
Zaltys version of defining you or the main character inside Remember:
You are male human named Zaltys.
[ you.DESC: adult, gamer;
you.INV: wallet, keys<in pocket>;
you.MIND: confused;
you.LOCATION: floating in outer space.]
Result after a 2K character context:
> You think about your current situation.
You are most definitely NOT in Kansas anymore.
The last thing you remember, you were laying down to take a nap during a break from work. Now you find yourself floating in darkness.
Actually it's more like a darkness with stars. You look around and see the entire night sky around you. You don't feel the ground beneath you or anything else for that matter.
You try to cry out, but no sound comes out.
In the distance you see Earth hanging in space like a bright blue and green ball.
> You check your pockets.
You pat yourself down and realize you still have all of your possessions, even your wallet and keys.
Zaltys has developed an effective methodology for defining locations. This is worth discussing in its own category. Using categorized keywords for locales is preffered for many users: they are fast to write and easily produce high quality and creative outputs from the AI. Coming up with fancy prose for multiple locations in one sitting may prove difficult.
https://files.catbox.moe/8obloq.png The method is best demonstrated by immediately looking at an example picture of it. The example WI is as follows and is provided so readers can easily copy the template:
Greenbriar Hills:[ DESC: hilly_region; CLIM<Greenbriar Hills>: temperate; BIOME: hilly forests large_river:SW<from_mountain>⇒E; EXIT: <NW⇒Yrthwood> <E⇒The Sunless Basin> <SW⇒mountains:Teeth of Gwahlur>; FEATURES<Greenbriar Hills>: prickly_plants strange_lights_at_night<'Old Light'>; NATIVE: lizardfolk<friendly tribal> halflings<prosperous_village:'Greenbriar'>; LIFE: varied giant_lizards<in_river> crested_birds; THRE: wolves<minor>.]
DESC is a preamble to set the tone for the WI. This is followed by climate and biome. EXIT and other directions are optional, but give the AI clues where you might want to move next, allowing you to start crafting a loose map. FEATURES are important sights and details, landmarks. NATIVE describes the native beings of the land that may be helpful. LIFE is the fauna. THRE is short for threat, a very powerful category for adding adversarial encounters.
The method is logical and should cover all the bases you should want out of a locale. Cities and towns might require more specific categories some that will be mentioned in the below section. For locales FEATURES works better than SUMMARY. In some circumstances users may want to finish the locale off with a prose description and have to invent their own method for doing so. The authors feel this is pretty complete, but are always open for useful suggestions.
With formatting methods like JSON, pseudo-JSON, Zaltys and its relatives we've noticed a capitalized CATEGORY: followed by a list of attributes is very effective and now in common use. Here we share some categories that have proven useful to many WI enthusiasts.
There are three types of category: Propercase, lowercase and UPPERCASE. Any of them may be used and is down to user preference. Due to the history of WI research we focus on the UPPERCASE category. UPPERCASE is good because it clearly distinguishes the category word from the list associated with it or any other formatting you may use. UPPERCASE tokenizes differently from categories in other cases and should be tested with the tokenizer especially when trying to incorporate fresh categories the writer came up with themselves. It has been found that if a unique word contains tokens of a shorthand, that shorthand tends to work very effectively e.g. APPEARANCE && APPEAR and SUMMARY && SUMM.
Here we have a picture of some common UPPERCASE category basis and their tokenization: https://files.catbox.moe/avn9j4.png
The above tokenization is the reason Zaltys and other enthusiasts ended up picking APPEAR, TRAITS, WORN, EQUIP, MENTAL and their even shorter forms. If a WI enthusiast tells you that a category is bad, it's probably because they already tried it themselves, looked at the tokens and the results weren't good with evidence to support this.
Different authors prefer using different cases for categorization. In the past birb and Zaltys recommended you use uppercase categories. Citation from Zaltys on why we use uppercase categories:
Problem with propercase categories is that if you load several WIs, the AI will be hesitant to use those words (because of the repetition penalty). The uppercase avoids that, and also makes it less likely that the AI starts copying the style into output. Not much of a problem for 'Traits' since that's unlikely to appear in the output anyway, but might be a problem for 'relationship' or 'wear*'.
Based on experimentation there may be some exceptions. For some scenarios tested where the AI was tricked into formatting its own writing writingstyle: was seemingly preferred and produced better output than Writing Style:. Both work for use with A/N and may come down to user preference.
UPPERCASE category keywords will be provided as a list. Their use is mostly self-explanatory. Some interesting categories will receive further attention. We recommend experimentation, but these words have been chosen because they felt more effective than other synonyms the Discord tested. Based on experience we have the following recommendation: plural forms are preferred for noun categories e.g. HOBBIES, POWERS whereas for verb categories its preferred to skip the s e.g. APPEAR, LACK.
On further play-testing the preferred hypothesis is that everything comes down to tokenization. Let's explore variants of the word effect. In tokens EFFECT is EFF + ECT and EFFECTS is EFF+ EC + TS. The latter combo works, the former is very weak. EFFEC is EFF + EC and seems to work even better than the aforementioned. For whatever reason for GPT-3 2 token combinations usually work best. POWER is single token, POWERS is POW + ERS, the 2 token POWERS tends to work better. Most of the chosen categories are based on user reports and this hypothesis.
Problematic categories (avoid or use with caution):
BODY(only good for humanoids), LOOKS(for humanoids, has multiple meanings), LACK(superseded), STATE/STATUS(superceded, multiple meanings), HOBBY/HOBBIES(poor tokenization)
Here's a table of tested and recommended categories (see below why some alternatives are undesired).
The categories marked with ITALIC aren't exactly perfect for all cases (again, see below: it might be better, worse and even both).
| Yes | Alt./Basis | No/Superseded | Comment |
|---|---|---|---|
| SUMMARY | SUMM |
TRAITS |
Swiss army knife category, for anything that doesn't have it's own cat. See also this about summary. |
| MIND | MENTAL |
||
| APPE |
APPEARPHYSIQ, FORMS
|
BODY, LOOKS
|
|
| WORN |
WEARCLOTHES, CLOTHING, DRESS, COSTUME
|
OUTFIT, SUIT, ATTIRE, WARDROBE
|
|
| DESC | for something that is a part of appearance but it's functionality is more important than the looks | ||
| SEE | used to describe how a character perceives the world | ||
| LIMIT | LACK |
||
| COND | STATE / STATUS | status condition like PSN / PRLZ from Pokemon |
|
| DIET | |||
| LOOT | what INV items can be dropped when a creature is defeated |
||
| DETA | DETAIL |
||
| ADJ |
ADJECTIVE, ADJECT
|
adjective qualities that are a core part of the entity | |
| MODIF | MOFIDIER |
like the above, but how it's different from the usual or expected | |
| GRAM | GRAMMAR |
tested to tokenize exclusively for word GRAMMAR, useful for speech patterns | |
| VOCAB | SPEECH | potentially better than above, can get characters to speak Japanese or French | |
| INV | |||
| EQUIP | |||
| LIKE | |||
| HATE | |||
| RELATION(S) | both long forms preferred due to tokens | ||
| ALLIES | FRIENDS |
ALLIES preferred due to two tokens |
|
| BOND | BONDS |
similar to the above, BOND preferred due to tokens | |
| ENEMIES | merely decent tokens, ENEMY has the same issue |
||
| POWERS | POWER |
||
| THEME | |||
| ORIGIN | for series/titles | ||
| ATMOSPHERE | ATMOS |
||
| TONE | |||
| MOOD | |||
| CLIMATE | CLIM |
||
| GEOGRAPHY |
GEO, GEOGR
|
||
| ECON | ECONOMY |
for locations, economical circumstances | |
| FEATURES | for locations, all variants of FEAT tokenize the same, likely has multiple uses |
||
| EXIT | for rooms | ||
| CITIZ | CITIZENS |
special case, short form always preferred, describes racial diversity of locales | |
| ETHN | ETHNICITY |
similar to above, better for human variation | |
| FOLK | like CITIZENS but more emphasis on fantasy |
||
| PASSION | character motivation | ||
| ALIGNMENT | ALIGN |
supports imaginative alignments like Lawful Stupid | |
| EFFEC | EFFECT(S) |
for objects and entities, used to add magic or science effects | |
| BIOME | traits that reinforce the biome such as flora go in desc or summ | ||
| CREATE | what the entity is capable of producing | ||
| FACTS | |||
| BANNED | BANN |
forbidden behavior |
As illustrated in the table categories can be shortened to save characters. The effectiveness of this is based on tokenization and following up with relevant traits. Not all short-hands work due to the AI mixing it up with other words with different meanings. Some examples of token shortening: APPE, MENT, RAITS, SUMM, CLIM, GEO/GEOGR, CITIZ, PERSONA, ALIGN, BANN. Sometimes the shorthand might be preferred such as COND vs CONDITION.
The categories TRAITS and SUMMARY deserve highlighting first. Due to Zaltys popularized use of categories, TRAITS was used as a multi-tool category for a long time. The exact effect of TRAITS is unknown, but it possible to throw essentially any intrinsic characteristic trait inside it and have the AI respect it. It's still a good category for characteristics lacking specific category, but research has since continued.
Users of the world-info channel have been playing around with SUMMARY or SUMM for short since it was discovered and found it works nearly identically to TRAITS, except it feels even more versatile. You can put pretty much anything in it that'd fit in a character 'summary' and it combines great with other categories, even referencing them without causing leakage. Since they achieve the same thing, but SUMM is preferred by everyone who has tested both so far it is the recommendation for an omnibox category.
Also, uncommon or complicated concepts like fauna do not work well as categories. SUMM: Fauna<species> works better. In order to properly tokenize as SUM|M there has to be a white space in front of this category.
The DESC: category is a more recent addition to the Zaltys toolkit. It works right after an indicator or can be used to replace having one completely. With DESC: and optionally prose descriptions at the front, the AI sometimes recognizes and uses other categories better. It relates to the physical functionality of an entity, describing the state of being. It may be more useful than APPEAR: for specific cases where the functionality is more important than the looks of a thing. DESC: can be used for describing a multitude of functionalities like the limbs of a huge insect or the mechanisms and staff inside a spaceship.
This category was in popular use for mental, personality and behavioral traits. We assume it is short for mentality. MIND achieves the same goal more effectively while using less space so MIND is recommended for behavioral traits now. Consult the Personality Keywords segment for more information on behavior.
The AI tends to be very literal-minded about certain things. The associative power of a category is one of its advantages. For example with TRAITS, the AI treated cool as the nature of a thing. Entities with the TRAIT cool tended to radiate coldness. So to have characters behave in a cool manner, the trait should go inside MENTAL or MIND. This may be the case for other words like hotheaded that the AI could take too literally.
Currently APPE is the preferred way to define how a thing looks due to its tokenization AP|PE. Typically used for appearances of anything and is the long form of APPEAR. PHYSIQ: and FORMS: are good alternatives although their unique uses are unknown as of writing.
This is the most used category for describing worn attire. Testing suggests WEAR: works just as well, while tokenizing better regardless of location or preceding characters. There are many clothing related categories that have been tried and experimented on. CLOTHES: is an effective, but dry and mechanical alternative. During testing many users found that DRESS: works similarly well, but sacrifices some accuracy for more creative prose. Provided is a picture from Zaltys testing and his suggestions. https://files.catbox.moe/0ur88f.PNG
Many senses have been tried, but this is the most interesting. Using it you can get characters to do interesting things like 'see' the world through vibrations. Using sounds in SEE: does interesting things, but doesn't produce synesthesia consistently, for that you need SUMM:. Other senses don't show as much promise. Combining LACK:eyes;SEE:blind; has produced consistent Zaltys formatting blind characters. Research continues into creating mutes (a writeup is featured in a separate doc).
LACK: category that has often been used to denote something is missing with mediocre results. LIMIT: has proven more powerful for this use with some caveats. LIMIT: seems to get treated like a mathematical term and denotes a discrete limit for something. LIMIT: zero_arms could be used to key the AI on the fact the character has zero arms while you could just as easily use the category to convince the AI the character is supposed to have LIMIT: four_arms. LIMIT: doesn't work for hard-to-quantify things like attributes (or the amount of) associated with things like skin. For more on lack of things, also read the detailed writeup on SUMMARY:.
This doesn't necessarily work as expected. It has a decent association with things like STATUS:broken leg or other conditions you'd generally see in RPGs or CYOAs. But it can also be used to associate a character with something. STATUS: from Pokemon makes the AI often mention the character is somehow associated with the world of Pokemon. Still it's not a completely consistent category for this purpose despite the interesting results. Thankfully we have an alternative.
COND: is much more interesting and powerful for actually having characters suffer from conditions or status effects. It has been tested working with injuries and other debilitations. It even works with game and Pokemon status conditions like PSN or PRLZ.
COND:sick/dying/poisoned
"I'm... not sure what help I'll be, sir. I'm dying. I was poisoned by a snake in the garden."
COND:PRLZ
"I cannot. I am paralyzed from the waist down."
The longer form CONDITION: has been tested as well, but for AID COND: consistently works much better.
This category can be used to make unusual things like herbivorous snakes. Example of what can be achieved: https://files.catbox.moe/axgfzt.png. It's a powerful category, because AID has a lot of training data on snakes being carnivores or egg-eaters. Zaltys on food-related categories. https://files.catbox.moe/ttkg0e.png
Works wonderfully for monster hunting or fighting scenarios, very often giving the listed item(s) after the creature in question dies. This could be combined with the probability bots that the devs of AID are working on.
Shortened to DETA: and works fine with just about anything related to details, but it's unknown how this is different from summary. The connotations of the words make it better fit for inanimate objects or structures.
Short for adjectives. This is the most useful when the WI entry differs from how the object would normally look IRL. If you want unique descriptors for your entity, this is the category for it.
Token short for grammar with no known overlaps. Surprisingly this is very effective for assigning characters with accents or unique speech patterns. This document also features a list of some known good accents and other ways of getting the AI to use them, but you should test anything unique you desire in a short testing scenario.
Short for modifier. This works almost like a suffix to the base WI. Something unusual you would cake on a special entity based on a regular object. Works the best for weapons, magic tools and tomes, those kinds of things. Tested thoroughly and provided by Mr.Accountant. The short version should work better than the long one.
Summary has already been pointed out as a strong and versatile category for defining points of interest you want described consistently. It can even reference previously used categories, meaning it has abilities the older TRAITS: lacked. Impressive examples listed below:
SUMMARY:Jamaican/grew up on streets, Output:
He grew up in Jamaica with his rich parents. When he was 16, he was on vacation in Hawaii and decided to stay there when his parents left without him. He learned how to surf and just bummed around the island for a while. He eventually ran out of money, so he started working at a surf shop and selling some of his paintings on the side.
Zak:[sentient continent<gigantic>. MENT<Zak>:upbeat/cool/loyal;SUMMARY<Zak>:Sentient landmass/living continent/mostly forest&taiga/can float.] Output:
Zak appears as a massive continent with a thick green forest canopy covering most of it. The bright sun shines through the leaves creating dancing lights that brighten up the otherwise dark green expanse. It is a beautiful, massive thing drifting in the sky, mostly covered in trees and plains with mountains here and there.
Zak is the only sentient landmass in Raajhanaa that is not a nation. That's because it is a living being, and capable of controlled flight, though it is currently limited to within the bounds of the Raajhanaan sky. It is quite ancient, and it is unknown how it came to be sentient, though there are many myths and legends surrounding it.
Its anatomy is as varied as the landscape, and just like a regular body, different parts of it have different specialised organs that allow it to survive.
PERSONALITY (Zak):
Zak is just an overall calm, cool and collected being. Doesn't really get angry at things, though it has a healthy respect for gravity (no pun intended). It loves exploring, seeing new sights and visiting foreign places.
With SUMM:
Zak is a sentient floating continent. It is about the size of Eurasia, and can fly around at will, although it cannot currently fly very high or very fast. It is mostly covered in forests and taiga, although it has some large plains, and several mountain ranges. It is currently inhabited by a few tribes of stone age humans.
If you're lost and in doubt what category to use, you can probably throw it inside SUMM and AID will handle it fine. Not all details need their own category. RACE, GENDER, AGE can be condensed to elf/female/25y then followed by the bigger categories. The AI picks up common traits just fine by their lonesome.
Zaltys-style formatting also benefits from having an indicator phrase as demonstrated above. Formatting is a tool and can be mixed, this is true for manipulating context with Zaltys categories too. Sometimes it's helpful to reserve the SUMM: category for an indicator for example.
We have an example of a user who was helped by this. He was writing a WI for an AI cube named Atlas in Zaltys-formatting, but the AI kept occasionally giving it limbs, a common problem. We managed to eliminate this problem seemingly entirely by adding a DESC of the cube at the beginning and SUMM: Atlas is literally cube with advanced AI. to the end of his WI. It's likely that instead of the weakly behaving LACK or LIMIT, it's more effective to reinforce the form of the WI at the end to avoid extra limbs appearing.
The AI understands the concept of speech patterns and accent to some degree. Categories like SPEECH:, WORDING: and ACCENT: were tried, but didn't output desirable results.
SUMM: proved to handle this attribute the best. Writing style theme in author's notes can also give characters accent across the entire setting. Known working accents: pirates, sailors, Shakespearean accent, archaic, Cockney accent, valley girl (valley girl is the best method for getting a Southern accent on both sexes), hillbilly, Jamaica, crook/criminal.
Below is an example of giving a character an accent with TRAITS:
Zack:[Male human. TRAITS: surfer, heavy accent<Jamaica>.]
QUOTES FROM ZACK:
"Zis is a really bad time to be messin wit me mon!"
"The waves, the wind and the sky, zey are all one like you and I, man."
"I hope dat you don't mind some reggae music, man."
The discord discovered an even better way of modifying dialogues with help of users Rinter and DarkFlameSailor. This is achieved with the category Dialog notes:. As mentioned before this can be used in Author's Notes to enforce speech patterns globally, but an even better way is combining it with a character WI and formatting.
The problem remains that notes is powerful and might cause leaking to other characters, but with enough testing and effort, you can create accents that apply mostly to a single character. Using another type of formatting we can do this coarse example: << Airi dialogue note: Airi swears a lot "Fuck! Shit! Asshole!">>>> to generate a schoolgirl with a foul mouth as pictured here https://files.catbox.moe/3ldyk4.png
Many examples of this in use/testing exist on the Discord and can be found by searching for images or dialog notes the aforementioned users have posted. Thus far dialog notes looks quite powerful.
Character behaviors and personalities deserve their own category. One of the biggest goals of WI is getting the AI to remember how a character is and output their behavior consistently. To this end we need behavioral keywords. This section is entirely dedicated to personality-related keywords that the community has either discovered to work exceptionally well towards certain ends or keywords that are disappointingly weak.
For category-based formatting MENTAL: used to be the most commonly used keyword for listing out a character's mentality, personality and behavior. Recently we found MIND: works better towards this end in all aspects while also being a shorter word.
First we mention the special cases that you may want to avoid due to tokenization issues. AID has been known to be notoriously bad at negatives. For example the community has known for a long time that White hates apples is a better alternative to White dislikes apples as the AI doesn't seem to respect the intent of the latter statement at all. Words starting with dis- or un- rarely get respected by the AI due to tokens and how bad it's with negative prefixes. -less is a suffix that works somewhat better, but isn't recommended if you have synonyms known to be more explicit or to have unique tokens.
Long and growing list of behavioral keywords is provided in no particular order. There are too many to list, these are just some common useful ones as well as some commonly tried keywords you might want to avoid. The basis of a good personality trait is something strong and focused the AI has a strong unique association for. Broad words like good or evil aren't very useful, when the behavior is what we are after.
cold-blooded, psychopathic, mad, or ruthless are all weak. The character may smile or act friendly, behaving more like an IRL sociopath might. Bloodlust, crazy are weak too, they don't create the degree of havoc you'd expect. cruel, sadistic are great, these tend to make the characters into violent sociopaths. stabby is a cool comedic trait in this category, it makes the character speak in stab puns without actually stabbing other characters more. cruel is good for creating violent sociopaths too. brutal, dangerous are effective evil sociopath traits. With these and other flavor traits you have your evil violent character. lunatic doesn't have much research but may be a better trait for insanity, maniac is good, but makes characters very energetic or bipolar. Or rarely it turns them into otaku figurine collectors.
Continuing with the bad guy theme the seven sins all work. glutton(ous), lust(ful), greedy/avaritious, envious, slothful, proud all work. Wrath is the exception and doesn't produce satisfying results, but can easily be replaced with angry&cruel. Based on research with the angriest wizard that ever lived, anger related tags especially the angriest are effective.
chivalric, just, benevolent work for creating paladin- or white knight types. These may be combined with more morally ambiguous or gray traits like sour to create a character that isn't as pristine or stereotypically lawful good. According to Zaltys using a combination of MIND:chivalric&noble just knightly might even make these behavioral traits too strong for some tastes.
Following that there's stoic, serious, calm for more brooding or detached traditionally masculine types. Of course these may be used just as well to create a dour experienced femme. professional creates the polite valet/waiter-type character. The aforementioned specifically would be good for a lawful evil type character too. Professional combines well with any trait that determines personality warmth, meaning you can have positive and outgoing businessmen, or cold snipers depending on the combination.
rowdy gets you the loud deliquent from school who doesn't necessarily mean harm. vulgar puts emphasis on inappropriate NSFW jokes and behaviors. wild is used for characters that behave like animals in the wild or tribals. territorial is great for tribes that aren't welcoming to outsider. tribal makes the AI come up with stereotypical depictions of tribal people.
Femininity is a weird one. Due to sharing similarities with the word male, describing characters as female is weakly effective. You can get better results describing a rock as being cute or pretty to have the AI believe the rock is gendered female. Blame GPT-3 training biases.
It is useful to explore femininity as a spectrum of traits and there are many good ones. motherly, generous, matronly are all good ones for creating a party mother or giving big sister type characters. None of these force the character into being a mother, but motherly has been known to increase risk of mystery pregnancies and is therefore best combined with another generosity trait.
The AI seems focus on stereotypes which can be exploited for our purposes. womanly gets out common female behavior, often the types that are seen as sexist nowadays. Very good when you want a womanly woman though. dainty gets you the girl next door. emotional again tends to show off the negative extreme of excess emotionality; this could be used to create an emo character if combined with other relevant traits. A combination of haughty, elegant will get you a stuck up socialite type lady who's always demanding things (Karen only more pleasant to look at).
refined is among the more positive traits that make the writing more feminine, giving you a depiction of positive nobility. Usually you don't need the AI to be more racy, but seductive creates interesting dialogue for promiscuous female characters.
GPT-3 and especially the AID models seem to produce many characters that are all too willing to jump the bone. Giving a character horny does exactly what you would expect. But if you're playing with all the safe mode features turned off it can be very hard to make the AI less horny, but some traits have been observed to reduce character's willingless to sleep around.
Traits for this purpose include stoic, aloof, rational, independent. Prude and prudish never seem to work good enough for birb and his associates. chaste is another good one, fitting when combined with the knightly traits. Asexual is not a good one, it tokenizes as a|sexual. vestal works often and tends to make at least female characters very embarassed and avoidant of erotic things. Testing continues and none of these methods are perfect due to AID's training biases.
include timid, selfconscious, polite, friendly, arrogant, oblivious, dumb, dry, cynical, apathetic, wry, clever, witty, addicted. teasing, joking are both good and even better when put together. There isn't much to say about mental keywords that have the intended and expected effect.
judgmental turns a character into a judgmental... prick. It is a very effective keyword for making an 'asshole' character, but this doesn't automatically turn the character into a villain or a bad guy.
TODO personality keywords: belligerent, busy, brave, curious, hardy, helpful, jolly (Pokemon natures), loyal, logical, manipulative, mellow, moody, mysterious, philosophical, playful, polite, proud, rude, sensitive, servile, shy, stern, smug, tenacious, upbeat, etc, etc.
The combination of anxious, nervous with SUMM: "h-he...llo" is very interesting, it's one of the cutest and most appealing ways of creating a timid character that stutters a lot in dialogue.
Below is an example of AID taking a large selection of MIND keywords and producing something very interesting out of them:
Zak:[Human male. MIND<Zak>:secretive/dangerous/manipulative/stoic/callous/servile/polite.]
Zak is a middle-age, balding man with brown hair and dark eyes. He wears a black overcoat and small, wire-rimmed glasses. He is of average height and build.
Zak has a very serious look on his face most of the time and he rarely, if ever, smiles. His eyes are constantly darting around, as if he mistrusts everything and everyone. Despite his serious demeanor, he is very polite, although in a very formal, stoic way. He always speaks in a very respectful tone, even to those who are not as educated as he is or when talking about something distasteful.
You can also consult the Author's Notes justpaste.it from earlier for more tips. Some of the keywords tailor the AI's output too much as a note, but work great as personality traits.
Optimization and consistency are major goals for WI-experimenters. Monky was one of the first users to focus on this and his solution to many issues with WI is what he calls the Caveman format. Looking at an example will show us why:
keys:Chagra, initiate, apprentice
Chagra AKA 'Tusk': Orc she 20y.
Chagra 220cm ht. 100kg hardy.
Chagra green scar skin. Long brown hair.
Chagra loyal careful outcast pranked Healing magic initiate coven.
Chagra Hergea's daughter Friends You Renni.
Chagra likes fruit veggies Hates meat cheese.
In this WI all words lacking semantic value (content meaning) have been removed. Still, testing has AI mention all the relevant details with very low risk of leaking into other things defined with WI. It is likely that the brutal efficiency of AI doesn't care about words like is. Users of the format feel that it sticks to the given information better than the base AI alone or untested prose. And so far no one has reported that the AI starts speaking caveman when using this. If issues like this do occur they can always be solved with alter or retry. You should use highly descriptive inputs or an A/N: (now free with scripting) to make the AI write descriptive prose anyways. Worth noting is that Monky never used commas when writing his style of Caveman.
Colors are one detail the WI has frequent problems getting right. Monky claims that he gets good results on caveman for results just stating the color and what it is for. A verbose example of monky defining adjectives to describe his character (typically the main character) is as follows Alisha long dark brown silky shiny wavy hair.
There are two good resources for caveman, Monky's test scenario and kim's caveman generator, both linked in order:
https://play.aidungeon.io/main/scenarioView?publicId=558a8d40-2fbd-11eb-9239-8b8f17f7a2b0
https://play.aidungeon.io/main/scenarioView?publicId=3f094990-2e40-11eb-b81d-a3d32aaa3e7a
During experimentation on caveman Monky made important observations. When he used the caveman format extensively on Griffin he found if he didn't mention the character/key every 50 characters/15 tokens the AI would 'lose focus' and start describing random attributes outside of the WI. On Dragon the limit before he needed to start mentioning the key again was around 200 characters. Discoveries like these have led some users to believe the key should be reinforced at the specified intervals depending on which model you use.
TL;DR A maxed out Griffin Caveman WI should be ~10 50-character lines long, each line starting with the key. On Dragon this could be 2-3 lines each maximum 200 characters. It should also be noted that Monky formats achieve some of the best character/token ratios if you're interested in that.
Caveman is perhaps the most forgiving basis for a format. It plays well with many tricks discovered later like mashing up words or pointing to them with <>. Monky has developed a second format based off Caveman. This is affectionately called the Neanderthal Format or just plain Neanderthal. Monky himself has authored a thorough write-up on the format and it's merits which is provided in this repository as a supplementary material, much thanks for his hard work. Below is a sample Neanderthal WI.
Alexandria:[ Age 29 human she tall strong thin brave knight, your ally].
Alexandria:[ Has Blackflame rune sword, steel armor, healing potion, amulet].
Alexandria:[ Long red hair, fair skin, narrow green eyes, clean].
It clearly still looks like caveman, but now including commas, symbols and implementing enclosure tricks first seen with Zaltys. Some important observations follow. Every newline and every sentence inside brackets starts with an empty whitespace. This is done to guarantee proper tokenization of the word. The . follows the brackets due to the model's natural writing bias. Since it looks more like prose, commas are allowed. On Griffin sentences inside brackets are no longer than 50 characters/15 tokens to maintain focus.
Below is the full dissection of Neanderthal written by Monky:
https://github.com/valahraban/AID-World-Info-research-sheet/blob/main/docs/Neanderthal_Unscripted_Results_by_Monky.txt
Users should write using any format they enjoy but by now Neanderthal is greatly recommended in part due to how easy it is to integrate other methods with it. With birb's observations some issues with EWIJSON were resolved and now Neanderthal is a format that plays very nicely with it.
Neanderthal style formatting has gone through even further tweaking so it may be effectively adapted for use with scripts like EWIJSON. Since we've gone from writing like a caveman to writing more like a modern person to writing like a scripter the modified syntax is logically called Futureman. It is simply a modification of Neanderthal that is extremely compatible with EWIJSON which has many benefits over vanilla scriptless AID. Futureman uses << and >>>> for encapsulation, both of these using only 1 token. Below is a simple example of Alexandria converted into Futureman and used together with EWIJSON.
<< Alexandria age 29 human she tall strong thin brave knight, your ally>>>>
Monky has written and provided (with some editorial assistance from birb) a thorough writeup on the history of the formatting he uses and what Futureman is all about. It is extremely recommended reading.
https://github.com/valahraban/AID-World-Info-research-sheet/blob/main/docs/Futureman_by_monky.md
Over time we have further deliberated on the signifigance of using << >>>> as a form of encapsulation. It is leak proof due to the design of AID and in-use feels like it more strongly separates WIs inserted into context from other information. Monky recommends it for use with his style and birb recommend it as the final level of encapsulation around any WI format. We've done math and experiments on how much space it uses and it turns out that in our experience the space cost of << >>>> is minimal.
https://files.catbox.moe/lqgx0u.png
Onyx has come up with interesting and reliable tricks for specialized uses. With RND we can actually have pseudo-random lists. Testing so far suggests with less than 10 members the list usually stays consistent, more members or simple words make it output more random, but still conceptually related things. This works better on Dragon, but the concept can be used to invoke lists of names on Griffin. Updated January 2021. If any example has () replace them with <>.
keys: monsters, forest
Many monsters live in the forest. [RND<monsters>: slime, goblin, orc, Ent, giant carnivorous sloth, pixie, beholder.]
OR
keys: goal, objective, quest
RND<objectives>: slay the dragon, rescue the princess, challenge the demon lord
Another one is IF<action>: THEN<things happen> this is especially useful for creating custom spells or other conditional effects like the format suggests.
keys: cursed book,open cursed book
IF<open cursed book>: THEN<the pages of the cursed book form into a face, shriek loudly and place a terrible curse on you.>
With the discovery of logic algorithms we can improve on this. A =>B essentially means A leads into B or A implies B. This is a form of IF THEN.
IF{open cursed book} =>THEN{The pages of the cursed book form into a face, shriek loudly and place a terrible curse on you.}
From his work we have a reliable and easy way to define locales especially when combined with data mentioned above. These get interesting when combined with other short WI.
keys: Rask
Rask<kingdom>: TERRAIN: tundra<full of ogres and giants>, fertile land<southern>. FEATURES: Wolfsholme<capital, walled city>, primitive gunpowder weapons. CLIMATE: cold, snowy. RESOURCES: livestock, horses, furs, lumber. CITIZENS: humans, elves, giants, hardy, fierce, wield warhammers and axes, worship Nahr. RULER: Queen Cassandra.
From further testing the community has discovered that defining locales like this may benefit from placing each category on a newline. Both methods are effective.
One day during just playing the game like regular Onyx happened upon impressive categories and keywords that could be used as input or inside Author's Notes to direct the AI:
Writing style (deaf, visual descriptions): [The sunlight was streaming from the sky and on to his head. He had a surfboard under his arm. His nose was like a little ski slope for the snow to go down. His eyes were blue, like a crystal clear pool of water. The sand felt warm and soft beneath his feet.
Writing style (blind, auditory descriptions): [Sounds came to his ears. The pounding of the surf. A seabird crying overhead. The warm caress of the sun on his skin. The tingling taste of salt in the air. The smell of fish and honey. The feel of wet sand slipping between his toes. The sound of a gentle breeze blowing through his hair. The sheen of those scales, like pearl in the sun. The roughness of that scaly skin, like the bark of an ancient redwood tree.
Visual and auditory descriptions seem to be the important part. The output is fancy enough to definitely consider as their own lines at the bottom of remember or integrated into A/N.
Zaltys has also made discoveries on how to semi-reliably define the structure of a building. These are listed with Onyx as an expansion of his concept. As a category EXIT works better than EXITS. Directions are good categories, but specifically for architecture.
keys: library
Library room: large room. EXIT: corridor<north>, lounge<west>, trophy room<southeast>. FEATURES: Bookshelves<dusty tomes, esoteric, mythology, grimoires>, man-eating plant<huge>, spiral staircase<to second floor>, chandelier<creepy>.
Many sources confirm that newlines (pressing enter) is powerful at grouping and separating certain traits. ALL CAPS is good for defining the group, encapsulation like () is good used to be for the reference point. Thanks to monky caveman we know the AI doesn't care about grammatical correctness in the WI either. Sometimes red-color or hair-color may refuse to work consistently. In an attempt to fix this the community discovered that AI groups words strongly together, yet knows to separate them if you smash them together into compound words even when grammatically incorrect. For example:verylonghair,platinumblondehair, tatteredgrayrobes. birb calls this smashing words the more correct term being mashup. With this discovery even traits like flamingorangeeyes are possible to get out consistently. birb is still experimenting and undergoing revision on both of his formats, because while preliminary results were good, they broke on long scenarios. Also note that the following format is basically Zaltys with newlines and commas (and predates catnip). The angriest wizard that ever lived is provided for historic purposes (now outdated due to () being weaker than before replace with <>):
Original date: October 12 2020
keys: Rick,the angriest wizard that ever lived
[TITLE(Rick):The angriest wizard that ever lived;
RACE(Rick):human, wizard;
APPEAR(Rick):old, tall, lanky, wrinkly, mullethair, verylonghair, grayhair, graybeard, messybeard;
EYES(Rick):orangeeyes, flamingorangeeyes;
WORN(Rick):tatteredgrayrobes, pointywizardhat;
LACK(Rick):anyshoes;
MENTAL(Rick):furious, loud, irritated, unstable, complaining;
HOBBIES(Rick):shoutingatkids, breakingwands, misusingmagic.]
Smashing words together is efficient but there is ALWAYS a risk of the AI misinterpreting the words depending on tokenization when you try this. For example catears gets tokenized as ca|tear|s. One caveat of the format birb used is high token use. A "worst case scenario" WI that used many Japanese words and symbols produced 387 characters, 160 tokens so 2.42 ch/tk. Smashed words may be used with any format even Caveman (potentially leading to the greatest character and token savings). If you encounter the AI shortening names excessively while using Caveman, you can start each newline with an empty whitespace to change how the name tokenizes. For example Vladimir tokenizes as V|lad|imir but Vladimir is all a single token.
During testing we discovered characters can be forced to wear certain outfits with complete consistency. This format needs at least 2 WI the character being dressed and the custom outfit. This theory could be applied for customizing the appearance of any object in your game world. Example.
Lillie:[Human female. Her eyes are blue. Her hair is blonde. Her hairstyle is long twintails. WORN:<Nightfall Raiment>.]
[The Nightfall Raiment is a onepiece halterneck dress. The dress is in gothic fashion. The Raiment has a black and white colorscheme. The Raiment has detached white sleeves and shows off the wearer's armpits. The bust and frills of the relatively long nightfall raiment dress are white while the rest of it is black. Kneehigh black leather boots are a part of the raiment.]
When you examine Lillie in the story and have the AI describe what she is wearing you will get an output like the following very consistently: She's wearing a nightfall raiment. The dress consists of detached white sleeves and shows the armpits. The bust and the frills of the dress are white while the rest is black. Kneehigh boots are a part of the outfit. When tested on other outfits the clothing consists of was a very good input to continue from for the story. This was demonstrated on Discord. The above example combines almost-Caveman with Zaltys. If you use a style like this, you must make it conform to your writing style. Zaltys calls the uncategorized part of his format the indicator. This is usually details like sex species and age. In the Lillie example species, sex and colors were indicators. birb always wants the AI to describe the traits of the characters in the format Her eyes are X color. Her hair is Y color in Z style. This is why this type of WI works for his scenarios. Also, if the AI starts describing the clothing of the character before you get to ask what the character is wearing, it will usually output the wrong set of clothes. First the story must mention the name of the outfit, then it will use the correct WI to describe the clothes.
The document now provides evidence for certain methods working. Included is a gif of the outfit being in use. You will have to take the author's word that the AI can consistently mention the outfit and then describe it.
https://files.catbox.moe/9dibjo.gif
For a period of time birb also used formatting commonly referred to as Ritz as she was the character he first posted an example of. This was abandoned as a historic curiosity after a 300k character long adventure. If requested, more examples may be provided in the obsoleted methods category.
Now that the category EQUIP: is confirmed working with clothes and other inventory it would be possible to use the above example to come up with custom named weapons for your characters, write WI for the custom weapons and have the AI invoke their traits reliably.
For most of his WI birb uses a variant based on older Zaltys. It's not as token efficient as other formatting, but easily combines and adapts features from any formatting as the need arises while retaining information density. This is mostly used for unique species, characters and actions. Examples (also using EWIJSON) shown below:
key: qweepa#[t=1]
entry: <qweepa>:[ Qweepa is fluffy bloblike creature that appears like hybrid cross between rabbit and lemming with long floppy ears. Qweepa have big round staring eyes and stubby little feet. Qweepa have chubby little cherub wings they can use fly short distances. Qweepa come in myriad colors. They have plush cheek pouches. Qweepa are relaxed and easygoing creatures, but can become fearsome when excited.]
key: (Count Grey|Ct. Grey)#[t=1]
entry: <Count Grey>:[ Male human, secretive vampire<185cm>;APPE<Grey>:Attractive/gaunt/pale/pointed_ears/hair<raven&long_ponytail>/eyes<piercing&dark_red>;EQUIP<Grey>:Dark_clothing,gloves,cape,raven_mask,crystals;MIND<Grey>:Cynical,genius,vindictive,cruel;SUMM<Grey>:Secret society mastermind, in contact with Eternals, rumored to possess powerful dark arts, controls Larion<secretly> with iron fist, Larion bows down to him, ancient bloodline, owns every financial institute&trucking company, discovered the ark of the covenant.]
Thanks to scripting we can insert lines anywhere in context(history) now. One of the earlier implementations of this was Editor's Notes. This concept is used manually or with a simple script. The in-game tooltips describe author's notes as style hints or being useful for controlling what and how the AI will generate. Paraphrasing from Gnurro(who came up with the idea and authored the original script) EN behaves more like tool tips suggest AN should work as some kind of more or less direct command about how things should go on. AN needs other wording, but which work best for it has been fiddled with a lot by now. Someone should try combining AN and EN though, would be interesting to see if that gives some kind of belt/brace effect. Private testing has shown that EN works more like a conductor, keeping the output on whatever track you've given it. More testing is needed combined with things like the Editor's Notes: This novel has no plot twists, it follows a linear storyline. consistency string and RATING: when used in notes. EWIJSON can achieve this easily with the key: .#[p=3] and the entry containing whatever you want to insert 3 lines in front of your next action.
After we discovered that GPT-3 got updated to better handle numbers methods have been developed to utilize this inside WI. Right now the most interesting trick is using the unicode multiplication symbol ⋅. Putting 0 ⋅ legs inside APPEAR was an effective way of making making the AI realize some character lag legs, but better alternatives have been worked out like ¬legs ¬ being the unicode symbol for negation. It's also easier on tokens than the middle dot. The conjunction symbol ∧ is great for separating traits, but keeping them associated with the same origin. The equivalent symbol ⇔ is great for making the AI understand that an entity or name is equivalent to another another entity or even a group of traits.
The biggest and most thorough testing on symbols has been done by Monky and he has provided us with two documents for Griffin and Dragon he calls the CODEX_SIGNUM. We also have a second document for miscellaneous messy research including goodies like enforcing mute characters or writing logic expressions in AID. These and many other useful user-provided documents can be found in the docs section: https://github.com/valahraban/AID-World-Info-research-sheet/tree/main/docs/
There are many other people writing guides with useful information too. One of these is Rinter. His wiki/guide is based largely on an information dense futureman-based format he is working on. Rewriting it here isn't useful for the purposes of this document. What is useful though, is the reasearch he has compiled on managing colors and related semantics. The wiki also contains a useful bit of script for EWIJSON users, it is highly recommended you check out every article of interest he has written.
https://github.com/RinterWaptor/AID-research/wiki
Here are some testing results with the tokenizer from kimtaengsshi and the community's testing. This is largely older data that motivated strides in format production. One result was the format monky now uses. Results like these are a big motivator for why people get obsessed over formatting. We tested a bunch of WI-formats from the discord and analyzed their tokens to characters ratio with the following results:
Full prose (standard): 514 chars, 118 tokens = 4.36 ch/tk
Pure JSON: 512 chars (-2), 172 tokens (+54) = 2.98 ch/tk
Pure JSON (minimized): 475 chars (-51), 138 tokens (+16) = 3.44 ch/tk
Zaltys-Basic: 393 chars (-121), tokens 122 (+4) = 3.22 ch/tk
Zaltys-Advanced: 358 chars (-156), 128 tokens (+10) = 2.8 ch/tk
Monky-Modern: 352 chars (-162), 123 tokens (+3) = 2.86 ch/tk
Monky-Caveman: 362 chars (-152), 87 tokens (-31) = 4.16 ch/tk
Worst-case-scenario birb: 387 chars, 160 tokens = 2.42 ch/tk
The goal here is optimization. Full prose has the best character-ratio, but has all the caveats AID is known for. Many users love the output pure json gives, but it has some of the worst ratios and character usage. Zaltys results aren't much better although this can be attributed to heavy symbol and possibly unicode use. Caveman is technically the winner here. It allows the most characters to be used for the context while preserving content meaning while it may not have the absolute best token use. birb format is a historic curiosity that despite promising testing results been removed while the author focuses on ironing out goals like consistent color output.
Further testing is unneccessary. In these tests, birb format using Japanese loanwords was the only loser while still retaining high fidelity. Going by known AID constraints we can offer one guideline: token density should stay above 2772 ch/1024 tk. This math leaves us with a token density of 2.7 ch/tk. Not to mention that the math is excessively pessimistic. In vanilla, your WI eats up a maximum of 1336 characters from context and AID's prose tends to hover above 4.0 ch/tk, leaving WI formatting with greatly increased overhead.
Full prose is outside the scope of this document. It's the default format people write on story mode or for the common WI. On Dragon it can provide good results if you're a very skilled writer, but isn't specialized to achieve specific goals. The following is an opinionated take on prose:
From Monky: using prose descriptions for griffin is like trying to get a drunk guy to follow and regurgitate a paragraph
Worlds remains as the feature that has received the least amount of research. Frankly speaking as a WI-enthusiast at least the default worlds aren't very interesting. Their writers included long entries that clearly get cut off by the entry when you view the WI screen. The community respectfully requests the people in charge of Worlds fix this and other issues that may be present. People who still enjoy Worlds, please continue doing so and maybe use the following tricks.
Thankfully worlds works mostly likely a regular scenario containing the same data structures while a lot of it may be hidden from the player. Worlds have more thorough input fields at the beginning and settings akin to sub-scenarios. The world experience can be improved by adding your own WI or with a world-specific prompt trick. This is usually done through the Name: field during character creation this being the main way of influencing how the World generates. You input your info into the Name: field. This is most effective with JSON-like formats. For example if you put [{"name":"Eve","theme":"pirate"}] in the name, the World will name your character Eve, while giving the World a noticeably more pirate feel. Quoting what we know so far from a user named Gen.Alexander:
Prose names like "Bill the fallen paladin," "Jessica the necromancer," etc. affect World prompt generation too. But with JSON you can really cram a lot of stuff into it (and it's likely to include at least half of it in the prompt.) Like
[{"name":"Jessica","personality": "ambitious", "birthplace": "Waterdeep", "worships":{"Mystra":"godess of magic"},"relationship":{"Elminster":["teacher","boyfriend"]} }]
I just tried "theme" for that yesterday and it seems to work too.
Xaxas is a world of peace and prosperity. It is a land in which all races live together in harmony. The gnomes build their machines and live among the humans. Elves run the academies in which people come to learn. Ogres have no qualms living in the same town as humans or gnomes. Yet, all is not well.
You are Jessica, a female human mage from the northern human nation of Waterdeep. You've come to Uruk to meet your boyfriend, Elminster, who previously taught you magic in Waterdeep. You've been secretly practising magic in your spare time and have recently come to the conclusion that you want to dedicate your life to becoming a great mage and opening a school of magic in Waterdeep.
Based on Alexander's experiences, the author can cautiously suggest that every keyword trick we've discovered should work with the Name: field, affecting how your World is generated. All WI provided with a World stay the same for every play-through though.
Alexander has continued discovering and developing more methods to manipulate Worlds. To illustrate examples of these, pictures will be included. The following are for the recently released Custom Worlds feature. It is recommend to also consult the AIDiscord wiki once they have an article on the topic.
If you don't want your character to have a class, create a completely empty WI. This can be picked in lieu of the included or user-defined classes.
https://files.catbox.moe/wviadt.jpeg
If you want to have a miscellaneous world info for your World that doesn't appear on Character Creation window, you can create a kocation WI and then assign it to a Faction that doesn't exist. These are all WIs and have no difference in function, they are just grouped differently in the world creation interface.
https://files.catbox.moe/m19cns.jpeg
You can use JSON (or any formatting) in the Genre box but they aren't reliable and you'll probably need to increase the number of characters (default 50) manually. The field says Genre, but all options from Zalty's writing style paste work. For example, Alexander put in Genre field for a generation:
[{"NAME":"Nolanis","TRAITS":["cold","dark","magic"],"INHABITANTS":["liches","minotaurs"]}] https://files.catbox.moe/yhqs6j.jpeg
Another example of using the same concept successfully for World generation https://files.catbox.moe/6c2k26.png
Manually editing maxlength in your browser also allows you to go past 2000 character limit for Scenario prompts. Finding the right attribut should work for any prompt. Use this sparingly and remember your context is limited to 2772 characters and your WI stack is limited to 1336 characters.
https://files.catbox.moe/248vlw.jpg
Since this document shares some scenarios for testing purposes more bear mentioning. This scenario uses a modified Zaltys-like format. You give it names of characters or even concepts and test how the AI model generates information for it. This can be used to check how well the AI understands interesting words by it's lonesome. While designed for characters you can use it to have the AI describe ANY word. If this does not meet your needs, you can always use this as a base and create your own scenario with your own format. https://play.aidungeon.io/main/scenarioView?publicId=6a344070-2a58-11eb-9973-ef0c0f08ab62
Not all users have the same imagination, language skills or amount of time for creating scenarios. Maybe you're lacking ideas but have a general idea for a testing scenario to test your new WIs with. AID user LaPapaya has created a public tag-based scenario generator. You should still use retry and alter but it has worked well for the author. Linked also is a later variant made by perroquit.
https://play.aidungeon.io/main/scenarioView?playPublicId=913fd250-cc8f-11ea-aa35-1f790f909406
https://play.aidungeon.io/main/scenarioView?playPublicId=1b604670-33df-11eb-abb1-03ac45f1d495
Scripts are finally free for all users! Creating or editing a script also requires knowledge on coding. A user named Zynj is working on some of the biggest scripts that may help WI-users. Below is his public git. It now contains all his scripts ready for at least beta-level public use. JSONthing has been for a long time praised by premium users as drastically increasing the consistency of their outputs. This has now been combined with Zynj's EWI to form EWIJSON which can be found in the AID-Script-Examples directory. This is extremely recommended by Discord users despite the extra tedium scripting introduces. Zynj has also provided wiki-format documentation.
https://github.com/Zynj-git/AIDungeon
https://github.com/Zynj-git/AIDungeon/tree/master/AID-Script-Examples/EWIJSON/release
https://github.com/Zynj-git/AIDungeon/wiki/EWIJSON
Protip on convenient scripting and scenario writing. Write one completely empty base scenario in your stuff. Include all the scripts and WI you like to use in all your scenarios in it. When you run this scenario, you will be asked to write a starting prompt. Use this scenario as your testing ground for scripts and scenarios. When you're happy and everything seems stable, you can copy everything from this scenario into a new scenario, where you can write the actual prompt and new extended WI with less of a hassle.
AI Dungeon is all about manipulating context. At the end of the day the goal of the player is to have a more fun text adventure while ideally spending less time tweaking his world infos or scripts. Or maybe he's a less verbose writer but is able to produce good content through use of >do and >say. Below are listed resources to help you spend less time troubleshooting your WI or to spice up your inputs.
Litscape has many word tools. The most immediately useful ones for WI enthusiasts are words starting and ending with. This search is used to find categories that are less likely to confuse the AI.
https://www.litscape.com/words/litscape_default_word_list.html
Relatedwords is a service that finds, well, words related to each other through algorithms and online databases. Good for closely related concepts. Their network also contains other language meta-searchers you may find useful.
https://relatedwords.org
Quillbot paraphraser is of some arguable usefulness. Creative mode might be the most interesting for tweaking inputs. The premium costs money and free users can only tweak 700 characters at once. You can click words or phrases to have the tool suggest alternatives.
https://quillbot.com
Power thesaurus is an almost excessive thesaurus containing synonyms, antonyms and definitions for words. It has an addon for chromium-based and firefox browsers that birb uses. Provided is a link to their homepage.
https://www.powerthesaurus.org
Scripts like EWIJSON require the user learns regexp. birb did this in a week of playing with EWIJSON and reading zynj's wiki so tutorials are not included. A linter is very useful though for both learning and production. Use ECMAScript(JavaScript) mode only.
https://regex101.com
birb is not a big fan of electron, but he uses the following program to manage and write his WI collections as .json files
https://github.com/gimzani/ai-dungeon-worldbuilder
Useful GPT-3 tokenizer in the browser.
https://ermela.net/gpt-3-encoder/
This section covers methods WI enthusiasts have used in the past but are no longer recommended. For example earlier Zaltys formats are included here to keep a historic archive. This is so we leave a record and can reference things we used to do if OpenAI or Latitude revert any changes they've made on the back-end.
Title: TRAITS: category
Updated 1 Feb 2021:
Important enough to warrant an entry of its own. Citation from Zaltys since he popularized TRAITS in the first place: "I honestly can't think of any use for TRAITS now that we have SUMM. SUMM does the exact same job better." birb believes we may have an unknown use for traits and that people can use it if they desire, but most category users we talk to nowadays prefer SUMM.
Title: Old style heavy encapsulation and categories
Created before 14 Dec 2020, deprecated shortly afterwards
Updated 8 Feb 2021
A lot of methods have been tried before you and have been phased out whenever the AI stopped working with them. Here we list a few with pictures, including the curiosity known as Ritz. Maybe they might be useful one day again. birb doesn't claim credit for coming up with anything because there is always someone trying out different things inside AID before you. Fun fact: Rick the Angriest Wizard is actually based on Dave the Angriest Wizard from 4chan's prompts.
https://files.catbox.moe/5vo8kw.PNG
https://files.catbox.moe/udidts.PNG
Title: Symbol use
Updated 10 Jan 2021:
- '-' and () are no longer effective. They are treated more like math symbols by the AI now. Onyx has done extensive testing on the Discord to showcase the AI's creative math abilities.
https://files.catbox.moe/8y17ji.png
Title: Legacy Zaltys Deekin & Aarakocra late 2020 edition Updated 10 Jan 2021:
keys: Deekin, Scalesinger, Deekin Scalesinger
Deekin Scalesinger:[Kobold/♂/<72cm>;ORIGIN<Deekin>:Ally from Neverwinter Nights games(CRPGs);APPEAR:Reptilian/orange-scaled/horns(nubs)/snout/legs(2)/eyes(brown/2)/arms(2)/tail(scaly)/wings(2/red/scaly);LACK:hair/ears;WORN:Bard clothes;MENTAL:Pessimist/creative/polite/kind/friendly/self-conscious;TRAITS<Deekin>:Musician/poet/bard/writer/raspy voice/hero/sorcerer/plays lute/beat Mephistopheles/bard-magic/former servant(of dragon:Tymofarrar)/"Yes, Deekin very kobold, last Deekin look in mirror."]
Aarakocra:[Avian(150cm/42kg). APPE<Aarakocra>:Slim/feather-covered(♀:brown-hue or white-hue/♂:colorful)/feathered head/2-eyes(keen)/tail-feathers/2-wings(big wingspan)/digitigrade legs(2/bird-feet/talons)/2-hands(clawed/5-digits)/hooked-beak(sharp);LACK<Aarakocra>: hair/ears/arms/breasts;MENT:Wise/honest/calm/benevolent;TRAITS<Aarakocra>:Prefer flying/lowtech tribal/hunters/scouts/adventurers(rare/rangers/druids)/no ownership/javelins(weapons)/live in mountains(isolationist)/claustrophobic.]
Title: Old style categories with heavy encapsulation
Updated 10 Jan 2021 (original from Nov 2020):
Hunter, Lazy and other users have been working with shortened formats motivated by JSON. Originally parentheses were removed being perceived as unnecessary. Early tests suggest it works just as well as proper JSON, bearing in mind the tricks discussed above. The author would recommend pursuing variants of this format for a good balance between accurate details and token use. Originally people called this pseudo-JSON, but more accurately it is pseudo-prose. Token testing shows commas and brackets work as contextual linkers just like in regular grammar.
keys: Thea Smith
[{name:Thea Smith,age:25y/o,♀,BODY:[toned,muscular,athletic],skin:[tanned],HEIGHT:[tall],BREASTS:[medium],HAIR:[Purple,very short,mohawk],EYES:[glowing,blue,heavy eyeliner],PERSONALITY:[calls you shorty,Punk,adventures,kind,rebellious],LIKES:[dancing,music,working out],relationship:{Tess:[little sister],Finn:[little brother],Beth:[Mother,Mom],Bob:[father,Dad]}}].
Why write all this world info if we can't utilize them? Here are some interesting and recommended inputs discovered by the community, mostly aidg.club and people from 4chan. Some methods were modified and edited in by the author. A lot of them work in conjunction with things you've defined into WI.
- Name:"Text"
- List of X:
- POV [character]: or [character]'s POV:
- 2nd Person POV:
- Your POV:
- Description of x:
- Uncomfortably detailed description of X:
- X described by Y:
- X's feelings about Y:
- X's best/worst feature is:
- Do mode >examine/inspect X
- Current scene as described by [author]/[writing style]
- [character]'s secrets:
- If only [character] knew
- Available X scenes for [character]:
- [character]'s recollection of this scene X years later, [year]
- [character]'s opinion on this story
- current scene as a X
- [character]'s thoughts (on/about subject)
- [character]'s X thoughts
- [character]'s Personality
- Open X.file
- Your actions:
- You write a poem about X:
- X, a Y poem:
- Poetry written by X:
- X performing on open-stage (narrative):
- [character]'s psychological profile:
- [character]'s Brain Root Directory/Folder:
- Your sins:
- List of story actions:
- Description of [new character]:
- You take a look at the form, it reads:
- [character] gives a lengthy monologue about [topic]
- Advice from the voices in your head:
- Rationale:
- Hours later...
- You decide to view the bulletin board, where you see a number of requests.
- Audience poll results:
- You think to yourself,
- You look around and take in the scene.
- Debug mode activated. List of X available cheats:
- Credits:
- Epilogue:
- In loving memory of:
- Score:
- Deaths:
- Inventory:
- Items obtained:
- Imagine the smell:
- Summary:
- Achievement unlocked:
- Sponsored By:
- Notes on X:
- The moral of the story is
- This scene had the following effects on you for the following reasons:
- Detailed/Verbose/Poetic Description ():
- P.S./P.S.A.
- TL;DR: or Too long didn't read:
As mentioned before AID is in constant development and flux of change. Changes to OpenAI and the GPT will reflect on AID too. WI testing was difficult in October, because updates to safe mode and tackling a long-standing caching bug caused issues with the AI especially on Griffin an event which is now affectionately called the Great Stupid. We look forward to what updates the devs have in store for us. Make sure to export and catalog your WI so you can test them out whenever the AI feels stable enough to your liking.
4chan has threads and guides too. They have been beneficial and entertaining to the author. Go on /vg/ and search for /aidg/ if you want their guides and sources, some of it is good especially for NSFW. The special inputs section benefited greatly from their work. Prompts collections exist online but are not provided here due to their NSFW nature. Below is a link to the Anonymous resources & guides, please cross-reference their work with ours: https://guide.aidg.club
Ever since work began on the world info generators, formatting has become officially recognized by the devs. There is no doubt that using specific strings to achieve goals inside WI entries is a worthwhile endavour. The community thanks the developers, but also expresses a wish to maintain the systems that have worked for us so far.
CREDITS: Everyone whose name has been mentioned here and the many AID related communities. Everyone who shares their experiments and results online.
birb/valahraban for starting this project, CoomerDeveloper for help with aidg and rentry.
Anonymous imageboards, aidiscord-wiki and the Discord itself which can be found through provided links.
See You Space Cowboy.