Reference Roles Leak Check

---
name: reference-roles-leak-check
description: >-
  Gives every reference image in a multi-reference generation exactly ONE job (who, where, pose, angle, style, light, palette,
  material, expression, wear, inspiration, layout), writes a "do not take" line for each picture, orders the jobs so the one most
  likely to be ignored comes first, and runs a leak check on the result before it is kept: is each picture's job visible, and is
  anything ELSE from it visible (its face, its objects, its colors, its words). Use whenever a prompt carries two or more
  reference images (Nano Banana, GPT Image, Seedream, Seedance reference-to-video, Lovart attachments), when a reference
  "leaks" (a stranger's face pasted onto a character, a style card's props walking into the scene, a palette ignored), when a
  reference's role is ignored, or when the user says "combine these references", "take only the light from this", "the face
  came from the wrong picture". Adapted from scenario-labs/skills (MIT).
user-invocable: true
---

# Reference Roles and Leak Check

A reference image lends EVERYTHING it shows unless the prompt says otherwise. Hand a model a film still as a style reference
and the actor in that still walks into your frame. That is how a pasted-on face appears on an invented character: a scene
reference carried its own person, and nothing told the model to leave that person out.

The fix: each picture gets exactly one job, and for every picture the prompt says what NOT to take from it.

Platform independent: works with any multi-reference image or video model. Reference caps and how the prompt addresses
images ("Image 1", `@Image1`, attachment order) come from the model's own docs; check them before numbering.

## 1. Assign one job per picture

| Job | The picture lends | Typical "do not" line |
| --- | --- | --- |
| WHO (main subject) | the character, product | "Use only this person's face and hair. Do not copy the sheet's layout, labels or background." |
| WHERE (stage) | the place | "Do not include any person from this photo." |
| POSE | the body pose | "Do not show the mannequin or any wood." |
| ANGLE | the camera | "Do not include the lighthouse." |
| STYLE | medium, render, texture | "Do not include the actor, the set or the props of this still." |
| LIGHT | the light only | "Do not include the flowers or vase." |
| PALETTE | the only allowed colors | "No green anywhere." |
| MATERIAL | a surface | "Do not show any fish, fins or water." |
| EXPRESSION | the mood of a face | "Only the mood moves. No skin, nose or hair from this face." |
| WEAR | clothing or armor | "The character's own face stays visible; no mask." |
| INSPIRATION | a design idea, not the thing | "A gown designed from this jellyfish. Do not show a jellyfish." |
| LAYOUT | grid and composition | "Do not copy the word MONDAY." |

Prompt skeleton, one line per picture:

```text
<One sentence naming the medium.> Each reference picture has exactly one job:
Image N (<what it shows>) = <JOB> only. <What it lends, concretely.> Do not <what to leave out>.
...
<Framing: shot size, placement, mood, text rules.>
```

## 2. Rules that held

- **One job per picture.** Two jobs on one card (subject plus palette) blur both. A card that stays in the call but has no job
  gets "Image 4 = no job: ignore it completely."
- **One identity per picture, one main subject per run.** Two people in one frame: each gets its own WHO card from its own
  locked sheet (see `character-identity-lock-setup`), named by position ("Image 1 = the man on the LEFT"). Never let a scene
  reference supply a face.
- **The job most likely to be ignored goes first**, usually STYLE, with "This is the most important job." A clay scene came
  back as chrome CG until the Style line moved to the top.
- **Lend concretely.** "Copy its 3D render look, soft materials, warm bounce light" beats "use its style". A palette job
  names each element's color ("the scarf is red, the coat is teal"); "palette only" barely recolors.
- **Settle conflicts between jobs.** Every picture carries more than its job: a style card brings its own colors, a subject
  photo its own fur. With a palette card, the style line adds "do not copy its colors". When that would recolor something the
  user cares about (their character, their product), ask before running.
- **Idea, not copy.** When a subject photo must not survive literally: "Use only the idea of a gecko. Do not copy the photo's
  pose, lighting, colors or realism."
- **Text** is quoted exactly with "No extra words, no duplicate text." Letter-perfect copy goes on top in the edit, not into
  the generation.
- **Clean cards.** Plain backgrounds, no readable signs or logos. Generate a missing card with its own text-to-image run, one
  idea per card.

## 3. Leak check before keeping

For EACH picture ask two questions: is its job visible, and is anything else from it visible?

- WHO: the face matches the identity sheet, not any other face in the call (run `independent-frame-critic` for faces).
- WHERE: no person, vehicle or sign from the location photo that the frame did not ask for.
- POSE and ANGLE: no mannequin, no wood, no prop from the sketch.
- STYLE: none of the style reference's own subject, set or objects.
- EXPRESSION: the mood moved, the face did not.
- LAYOUT: the grid moved, the reference's own words did not.
- PALETTE: no color outside the swatch.

Fix a leak by tightening THAT picture's line (add its "do not", or move its job up) and running again, never by retouching.
Log the fix in `generation-failure-fixes` when it is new.

Before rewriting a line for an ignored job, confirm the model actually received that picture (check the request log or the
platform's run details). A reference dropped by the API returns no job, whatever the wording.

## 4. Into video

When the kept frame is animated, the video prompt carries no jobs, only motion, camera and sound: the frame already holds
every job. For reference-to-video (Seedance with several images) keep the same one-job lines and add "Almost static, these do
not change: ..." for anything that must hold (see `seedance-prompt-writer`, `video-clip-continuity-chain`).

## Common mistakes

- **The main subject copied the photo.** Put STYLE first, use "only the idea of", make the palette strict.
- **A face from the wrong card.** A location or style still with a person in it, and no "do not include any person" line.
- **A logo by accident.** Generated billboards in red, white and blue swooshes read as a soda brand and get refused later. Say
  what signs show (waves, clouds, cranes), or "invented, blank signs, no logos".
- **A protected design named in the prompt.** Describe an original design instead of naming it.
- **Drift in the film.** A guitar changed color mid-clip. Ask for "almost static", list what must not change, move only what
  should.

Related: `storyboard-reference-assembly`, `segment-reference-lock`, `style-lock-guard`, `identity-fix-pass`.