The moment it goes wrong

You have forty shots back. Thirty-four are beautiful. Then:

• Shot 9, her face is someone else's face
• Shot 14, the coat went from navy to black
• Shot 22, the light is coming from the other side of the room
• Shot 31, she aged eleven years

Nothing is broken. Each generated correctly from the prompt it was given. The prompts drifted, one edit at a time, over an afternoon.

The fix is not a better prompt

It is one file that the prompts are built from rather than typed into.

Then shot 40 and shot 1 cannot disagree, because neither was written by hand.

What goes in it

One folder, plain text, readable by you and by it:

bible/
  characters/elena.md      face, build, hair, wardrobe, the non-negotiables
  characters/marcus.md
  locations/apartment.md   the room, its light, its era, its mess
  look.md                  lens, grade, grain, aspect, the whole film's rules
  refs/elena-01.png        the approved reference frames themselves

Text, not screenshots of text. A file it can read is a file it can apply.

What a character file actually says

Specific, physical, and short enough that it survives being pasted into every prompt:

Elena. Late thirties. Dark brown hair, shoulder length, always slightly unkempt, never styled. Heavy navy wool coat, three buttons, collar up. Small vertical scar through the left eyebrow. Tired around the eyes in every shot, including the happy ones.

Every clause is a thing a model can execute. Compare "a beautiful mysterious woman," which it has to invent, differently, forty times.

The look file is the one people skip

Characters are obvious. The film's own rules are the ones that drift silently:

Nobody notices a consistent look. Everybody notices an inconsistent one.

Assembling instead of typing

Now the prompt for any shot is not written. It is built:

"For every shot in shots.json, assemble the image prompt from three pieces: the shot's own action line, plus the character file for anyone in cast, plus look.md. Write the assembled prompt back to a prompt field on the shot. Do not paraphrase the bible, use its words."

Do not paraphrase the bible, use its words. That clause is the module. A model that is allowed to rephrase her coat will rephrase her coat.

Change her coat in one place

The client wants the coat grey.

"Change the coat in bible/characters/elena.md to heavy grey wool. Rebuild every prompt that features her, then list the shots that need regenerating."

One edit, forty correct prompts, and a list of exactly what has to be re-run. Not a search and replace across forty prompt strings, which is the version where you miss three.

Seeds and references are the other half

A locked prompt narrows the drift. It does not close it. Two mechanisms actually hold a face:

• Reference images, the approved frame handed back in as the thing to match
• Seeds, so the same prompt can produce the same result twice

Any course that hands you a vendor's parameter is wrong by the time you watch it. The durable skill is making it go and check.

It still cannot see

It cannot tell you her face drifted. It has no eyes. Anyone selling you automated per-frame consistency checking is selling you something you should test before you believe.

What it can do is everything around the looking:

• Every shot that has her in cast but is missing the reference image
• Every prompt that no longer matches the current bible
• Every shot generated before the last bible edit
• The contact sheet, in shot order, so your eyes do the one job only they can do

The drift audit

This is the thing to run before you commit to a cut:

"Audit consistency: list every shot whose stored prompt does not match what the bible would generate today, every shot missing a seed or a reference, and every shot generated before elena.md was last edited. Sort by shot number. Do not regenerate anything."

Fixing drift without remaking the film

Thirty-four shots are fine. You do not have a forty-shot problem, you have a six-shot problem.

"Regenerate only shots 9, 14, 22 and 31, from the current bible, keeping every existing take. File the new ones as the next version and leave v1 alone."

Six regenerations, nothing overwritten, and the good shots never at risk.

The exercise

Your shots from Module 4, and about twenty minutes.

1. *"Read shots.json and draft me a character file for everyone in cast, from how they are

described across all the prompts. Mark anywhere the descriptions already contradict each other."*

1. Read the contradictions. That list is your drift, already measured.
2. Fix the files by hand. These are yours and this is the creative work.
3. "Write look.md from the prompts too, same way, and flag the disagreements."
4. *"Now rebuild every prompt from the bible, using its exact words, and show me which shots would

change."*

1. "Audit consistency and sort by shot number. Do not regenerate anything."
2. Regenerate only what the audit found.

Step 2 is the one that pays for the module. Most people have never seen their own drift listed.

What you should have now

• bible/ holding every character, location and the look, in words a model can execute
• Prompts that are assembled, not typed, so shot 40 cannot disagree with shot 1
• A record of seed, reference and settings for every frame you own
• An audit you can run before any cut, that reports and changes nothing
• Six regenerations instead of forty

Your film looks like one film, and the next one starts from this folder instead of from nothing.