Build a Character Bible for a Recurring AI Series
A character bible is a single written record of every decision you are not allowed to re-make: the approved references, the wardrobe states, the location anchors, the light direction, and the models each was approved on. It exists because by episode four you will not remember what you decided in episode one, and re-deciding is what breaks a series.
A one-off short is survivable from memory. You hold the whole thing in your head for a week and ship it.
A series is different. Three weeks pass between episodes, you change a model, you tweak a reference because a shot looked better, and by episode five the lead has quietly become someone else. The bible is the boring document that prevents it.
What actually goes in a character bible?
Five things: identity references, wardrobe states, location states, the look rules, and a decision log. Everything else is optional and most of it is procrastination.
| Section | What it records | Why it earns its place |
|---|---|---|
| Identity | Approved reference views per character, with file paths and the date approved | The single thing a viewer notices when it slips |
| Wardrobe | Each approved outfit as its own named state, and which episodes use it | Costume changes read as errors unless they are deliberate and tracked |
| Locations | Approved empty-room reference per recurring space, plus light direction and screen direction | Scenes fall apart on the room more often than on the face |
| Look | Lens language, colour, grain, aspect ratio, typical shot lengths | Keeps episode six feeling like episode one |
| Decision log | What changed, when, why, and from which episode it applies | The only section that gets more valuable over time |
Why does every entry need a file path?
Because a description is not a reference. Writing down rust overshirt tells you what you chose; it does not let you reproduce it. The bible has to point at the actual approved image.
This is the single most common way a bible fails. It gets written as prose, reads beautifully, and turns out to be useless the first time you need to regenerate a shot, because prose regenerates a new garment rather than the one you approved.
- Path. Where the approved file actually lives, not where you think you saved it.
- Date approved. So you can tell which version a given episode used.
- Model it was approved on. A character reusable across models still behaves differently on each one.
- One line of intent. Why this was chosen, so future you does not undo it by accident.
- Episodes it applies to. From episode three onward, rather than just active.
When should you stop improving the cast?
At the end of episode one. After that, every improvement is a continuity error dressed as a refinement, because it makes the character inconsistent with everything you have already published.
This is genuinely hard, because your taste improves faster than your archive. By episode four you will be able to make a better version of the lead than the one you shipped. Making it is the mistake.
- 1
Spend the effort before episode one ships
Design, validate against your hardest planned framing, and build the reference set at the scales you intend to shoot. This is the moment to be picky.
- 2
Declare the freeze in writing
Date it in the decision log. A freeze you did not write down is a preference, and preferences erode.
- 3
Collect wants instead of applying them
Keep a list of improvements you are not making. Most stop looking important within a month.
- 4
If you must change, change on a season boundary
A visible change between seasons reads as a production decision. The same change between episode four and five reads as a mistake.
The mechanics of building that reference set, including how to validate it before you rely on it, are in building an AI character reference sheet that survives new shots.
How do you actually keep it up to date?
Update it at the end of each episode, not at the start of the next one. Ten minutes while the decisions are fresh beats an hour of archaeology three weeks later.
- 1
Close out the episode in the log
What changed, what got approved, what broke. Three lines is plenty.
- 2
File any new approved references immediately
New wardrobe state, new location, new character. If it is not filed with a path and a date, it does not exist.
- 3
Note what failed
The shots that would not hold are as useful as the ones that did. They tell you what not to attempt next episode.
- 4
Re-read the look section before you start the next one
Two minutes. It is the cheapest continuity insurance available.
Frequently asked questions
What is an AI character bible?
A written record of the production decisions a series is not allowed to re-make: approved reference files, wardrobe and location states, look rules, and a dated log of changes. It is a continuity tool rather than a backstory document.
How is it different from a normal series bible?
A traditional bible records story and character. This one records artefacts. Every entry points at an approved file with a date and the model it was approved on, because that is what you need to reproduce a shot months later.
When should I stop changing my main character's design?
At the end of the first episode. After anything is published, an improvement to the design makes the new work inconsistent with the old. Save wanted changes for a season boundary where the shift reads as deliberate.
Do I need to record which model I used?
Yes. A character reference is reusable across models, but results still vary between them, so a reference approved on one model is not automatically approved on another. Record it and you avoid re-testing from scratch.
How long does a character bible take to maintain?
About ten minutes at the end of each episode if you do it while the decisions are fresh. It becomes an hour of guesswork if you leave it until you need it.
Episode six should look like episode one
Design a cast once, reuse it across every model on one credit wallet, and keep the record that makes a series hold together.
Consistent character AI