2.3 The Character Bible: A Complete Specification

The character bible is the document that makes a persona reproducible. Not reproducible as an image — that is the LoRA's job — but reproducible as a character, across months, across your own mood swings, and across the moment when you cannot remember what she said about her sister in March.

Six sections. Fill all of them before the first pixel, because every one of them is cheaper to decide now than to invent retroactively under comment pressure.

Section one: identity

The facts a person would know about themselves.

The rule for this section: write down only what you will keep consistent, and be willing to keep everything you write down. A blank field is safe. An invented detail is a promise.

Section two: physical

This is where the work of Parts III and IV gets its specification.

• Height and build, with the head-to-body ratio your dataset's full-body frame establishes (4.4)
• Distinguishing marks — a mole, a scar, an asymmetric feature. Counter-intuitively, adding one helps identity: it gives the audience a fixed landmark and gives your QA a checkpoint
• Hair: length, color, texture, and the styles the persona actually wears. Hair changes are the cheapest visible evolution a persona can have (2.4)
• Face: the twelve or so attributes the reference plate represents. Not the prompt — the plate
• Hands: nails, jewellery, any ring. Hands are the most reliable tell in synthetic imagery (1.4), and a fixed hand specification makes them checkable
• Voice: register, pace, accent, verbal tics (6.2). For a persona who speaks, this field is as important as the face

Section three: style

Wardrobe and presentation, specified as a system rather than as a set of outfits.

The palette. Three dominant colors and two accents, chosen to look coherent in a grid. This belongs with the brand kit in 2.5; the character bible states what the persona chooses, and the kit states how it is rendered consistently.

Brand posture. Fictional labels, or real ones? Real brands raise rights questions (11.4) and date your content; fictional labels are safer and less legible. A common middle path: unbranded, recognizable categories of clothing (an oversize blazer, a specific denim cut) rather than named labels.

Silhouettes and occasions. What she wears for work, for a night out, for a Sunday. A persona with four defined occasions generates content far more easily than one with a single aesthetic.

Seasonality. Northern or southern hemisphere, and what changes in each season. This field quietly generates a year of content variation and prevents the "same outfit in every post" convergence.

Section four: the world (the prop closet)

Everything around the persona. This section is the most under-rated part of the bible, because it is what makes content ideas cheap.

• Home: type, size, rooms, the recurring objects. A specific apartment is a reusable set
• Neighbourhood: walkable or not, a recurring cafe, a park, a gym
• Work: the place, the desk, the colleagues, the hours
• Transport: does she drive, cycle, take transit
• Pets: the single highest-engagement prop available, and a permanent continuity obligation
• Objects with meaning: a mug, a jacket, a plant, a book. Recurring objects are how an audience recognizes a world

The test for this section: can you list twenty places the persona can be, each with three things in it? If yes, you have a content engine. If no, every post is a new invention, and surviving chapter 2.6's cadence becomes a creative emergency each week.

Section five: voice

The section that decides whether the persona reads as a person.

• Sentence length and rhythm. Short and clipped, or long and flowing. Pick one and hold it
• Vocabulary: words she uses, and a shorter list of words she never uses
• Register: formal, casual, ironic, earnest. Constant across platforms, adapted per format
• Emoji and punctuation habits. Specific, consistent conventions. Overuse reads as marketing
• Opinions she holds, at least five, specific enough to be arguable
• What she never says. The most useful field in the whole bible. A persona defined partly by refusal is more believable than one defined only by preference, and the list prevents the slip that ends a character's credibility
• How she handles being wrong. Defensive, self-deprecating, curious. This is the field that governs every future crisis (8.6)

Section six: the canon

Ten facts that cannot be broken in any post, comment, or reply. The canon is the operational core of the bible.

Example shape — the specifics are yours:

1. She has never left her home country.
2. Her sister's name is fixed and appears in at most one post in twenty.
3. She does not drink alcohol.
4. She has never been in a relationship that the audience has seen.
5. She does not discuss politics or religion, and has a stated reason.
6. Her job is specific and she does not change careers.
7. She is terrible at one thing, on purpose, and it recurs.
8. She does not claim expertise she has not demonstrated.
9. She never contradicts a previous post; if she must, it is an in-character correction.
10. She is transparent, in character, about being generated, per chapter 11.3.

Number ten deserves emphasis. The canon should include the disclosure, as part of the character rather than as a legal overlay. A persona who is matter-of-fact about being synthetic is more coherent than one who avoids the subject, and it converts the largest structural risk in the category into a character trait. Chapter 11.3 supplies the wording.

How to keep it: a file, not a document

The bible is only useful if it is reachable at the moment of production. Three practical decisions.

Format: split it in two. Keep a short machine-readable core and a longer prose document. The machine-readable part is the fields a script can check — the canon list, the palette, the prohibited words, the recurring character names. The prose part is the voice, the backstory, and the reasoning. A single long document gets read once and referenced never; a small structured file gets read by your pipeline every day.

Location: in the persona's folder, with the dataset. Chapter 3.6's tree already defines the place — 00-bible/. It belongs beside the references and the training artifacts, for the same reason: they are one artifact set describing one identity, and separating them is how a persona becomes unreproducible.

Versioning: date every change, and never overwrite. A canon edit is a change to the character, and changes have consequences for what has already been published. Keep the previous version. When someone in the comments quotes a post from March that contradicts your new canon, you will want to know that the contradiction was created in August and not in March.

Update cadence: quarterly, deliberately. Chapter 12.2's audit includes a canon review. Most quarters produce no change, and the review is what prevents accidental drift — the accumulation of small improvisations that never got written down.

Three worked canon entries, annotated

The ten facts are the character's constitution. What makes an entry useful is that it constrains something.

Weak entry: "She loves coffee." Nothing can contradict this. It protects no decision, and it will not prevent a single mistake.

Useful entry: "She drinks black coffee and has an opinion about it; she never orders a latte, and never mentions a specific chain." Now the entry resolves questions: what does she hold in a sponsored post? Can a brand partnership be a coffee brand? What happens if a commenter asks her favourite drink? Three future decisions made in advance.

Weak entry: "She is honest." A trait, not a constraint.

Useful entry: "If she is wrong about something she said publicly, the next post corrects it in character within a week; she never deletes a post." This one governs every future crisis. Chapter 8.6's survival section depends on this field existing before it is needed, because the moment of a public error is not the moment to invent a policy.

Weak entry: "She posts about her life." Useful entry: "Her sister appears at most once in twenty posts and is never named in a caption, only in replies." Specific, countable, and checkable by reading the last twenty captions. That is the standard: if you cannot check it in ten seconds, it is not a canon fact.

The pattern across all three pairs: a canon entry is useful in proportion to how many decisions it resolves and how easily it can be verified. Write ten entries that meet both tests and you have removed most of the improvisation from a year of content.

The canon test

Before publishing anything: does this image, caption, or reply contradict the canon?

That test is why the bible exists. Without it, continuity is your memory's job, and memory degrades at exactly the rate your library grows.

Run it as a checklist item in the publishing flow from 7.1. One question, ten seconds, and it prevents the failure mode 2.4 describes — an audience catching a contradiction and documenting it in the comments.

One more artifact to keep beside the bible: a decision log. Whenever a post raises a question the bible does not answer — what does she drink, how does she refer to her job, which of two plausible replies is in character — write the answer down with the date. The bible is what you designed; the decision log is what you discovered, and after a year the second document is the one that keeps the first honest.

Where this goes wrong

Writing a bible and never opening it. It is a reference, and references that are not consulted are decoration. Put the canon in the publishing checklist.

Inventing details you will not maintain. Every unspecified field is safer than a specified one you forget.

A voice section with adjectives and no examples. "Witty but warm" is not a specification. Sentence length, vocabulary, and refused words are.

No prop closet. You have built a character with nowhere to stand.

Excluding the disclosure from the canon. It becomes an interruption instead of a trait.

Canon facts that constrain nothing. "She is kind" cannot be contradicted, so it protects nothing. Canon facts should be specific enough to be falsifiable.

The gate

1. Are all six sections filled, with every field you intend to keep?
2. Can you list twenty places and three objects each?
3. Does the voice section contain rules rather than adjectives?
4. Are there ten canon facts, and is at least one of them falsifiable by a single post?
5. How does the persona disclose being generated, in character?

What you leave with

Six sections that together make a character reproducible, a physical specification that feeds the reference plate and the dataset, a style system with seasons, a prop closet that turns content generation from invention into retrieval, a voice defined by rules and refusals, and a ten-fact canon whose consistency is checked by one question in your publishing flow.