Image Prompt Ambiguity & Drift Check

How it works

Image Prompt Ambiguity & Drift Check is for the moment before you spend generation credits. You paste one prompt or a batch, either a plain text-to-image prompt or an image-to-image edit of a reference, and it audits the wording the way a literal-minded generator might read it. It makes no API call, generates nothing and never claims what any specific model will do; it only says what a word "commonly renders as".

Phase 1 — Intake. Whether you gave a reference description (Mode R) or only text (Mode T; a mixed batch gets a mode per prompt), the numbered prompts, the task type and the intent of each, plus the tool's limits: it predicts nothing about a specific model.

Phase 2 — Five checks per prompt. A1 words with more than one visual meaning (glasses, bat, crane, mouse, plant, stamp, table, bank). A2 reference attributes the prompt never states and that commonly drift in derivation: list every attribute that must survive, then check each is stated. A3 contradictions, negations and overloaded prompts. A4 areas generators are commonly weak at (legible text, exact finger counts, asymmetric gaze). A5 wording that commonly trips safety filters. Every issue quotes your exact words, and the summary line counts issue-table rows per check (an empty check reads "none found").

Phase 3 — Rewrite. A safer rewritten prompt per input, keeping your subject, mood and style, with a one-line note of what changed.

Phase 4 — Attribute-lock checklist. For a single derivation the lock phrases go straight into the rewrite; for a set of variants, one shared table of what must survive, which prompt forgot it, and the lock phrase to paste into every prompt.

Sample output

Real, unedited output from running this Skill on a sample problem:

Mode: per-prompt (T/R) — 3 prompts — issues: A1 7 / A2 7 / A3 6 / A4 2 / A5 1

## Intake

- Mode: P1 = T (no reference given), P2 = R (reference described attribute by attribute), P3 = T (no reference given).
- Prompts:
  - P1: a cozy scene with a girl, two animals and a sign.
  - P2: re-places the reference woman at a beach.
  - P3: a stamp design with a mouse, in a mixed style.
- Task type: P1 text-to-image. P2 image-to-image edit of a reference. P3 text-to-image (assumed, no reference mentioned).
- Reference for P2, as described: adult woman, shoulder-length chestnut hair, round gold glasses, green cardigan, cafe background. Face shape, eyebrows, eye colour, skin tone, lighting, framing, art style and aspect ratio were not described, so they were not seen.
- Intent:
  - P1: a girl in eyeglasses at a table with a bat and a crane, cozy mood, a sign reading "OPEN", a hand showing six fingers.
  - P2: the same woman, same look, moved from the cafe to a beach, smiling.
  - P3: a stamp picturing a mouse on a plant, minimalist, with a night and a day version.

## P1 — cozy scene with girl, animals, sign

### Issues

| # | Check | Quoted words | Why it may go wrong ("commonly renders as") | Fix |
|---|---|---|---|---|
| 1 | A1 | "glasses" | Eyewear or drinking glasses; with "at a table" a drinking glass on the table is a common reading. | "round-framed eyeglasses" |
| 2 | A1 | "bat" | Flying animal or baseball/cricket bat; next to a "table" a sports bat is a plausible picture. | "a bat (the flying animal)" |
| 3 | A1 | "crane" | Bird or construction machine; "a bat and a crane" gives thin context and the machine is a common reading. | "a crane (the long-legged bird)" |
| 4 | A1 | "table" | Furniture or data grid; furniture is likely, low risk. | "wooden table" |
| 5 | A3 | "a girl with glasses sitting at a table with a bat and a crane, cozy, epic, beautiful, with the text 'OPEN' on the sign, 6 fingers visible" | Overloaded: about 11 independent attributes (objects glasses, table, bat, crane, sign; pose sitting; style cozy, epic, beautiful; text; finger count). Details commonly leak between subjects. | Split into P1a scene, P1b sign text, P1c hand. |
| 6 | A3 | "cozy, epic, beautiful" | "cozy" (small, warm) and "epic" (grand scale) commonly pull in opposite directions; "beautiful, epic" carry little direction. | Keep "cozy" and give "epic" a concrete form such as a grand backdrop (assumed). |
| 7 | A4 | "text 'OPEN' on the sign" | Sign text commonly comes out garbled, even one short word. | Generate a blank sign and add "OPEN" in an editor. |
| 8 | A4 | "6 fingers visible" | Exact finger counts are commonly off, and hands commonly deform. | Generate the hand separately and plan a check pass. |
| — | A5 | none found | — | — |

A2 not applicable: no reference.

### Rewritten prompt

```
P1a: A girl wearing round-framed eyeglasses sitting at a wooden table with a bat (the flying animal) and a crane (the long-legged bird) (assumed, change if wrong), cozy warm mood with a grand, epic backdrop (assumed), a blank sign on the wall.
P1b: add the single word "OPEN" to the blank sign in an editor.
P1c: generate separately, one close-up hand with six fingers visible (assumed, change if wrong), plan a check pass.
```

Changed: resolved glasses, bat, crane and table; split the overloaded prompt; replaced the sign text with a blank sign plus an editor step; made "epic" concrete; isolated the finger count.

## P2 — same woman at the beach

### Issues

| # | Check | Quoted words | Why it may go wrong ("commonly renders as") | Fix |
|---|---|---|---|---|
| 1 | A2 | "Same woman now at the beach" | "Same woman" is too weak to lock anything; a change of place commonly drags the whole picture along. | Say "keep ... unchanged, only change ..." (lock block below). |
| 2 | A2 | (not stated) hair; reference: shoulder-length chestnut hair | Unlocked. Hair colour and length commonly drift, more so with beach wind or sun. | "shoulder-length chestnut hair, same style" |
| 3 | A2 | (not stated) glasses; reference: round gold glasses | Unlocked. Accessories are commonly dropped or changed in a bigger scene change. | "wearing the same round gold glasses" |
| 4 | A2 | (not stated) outfit; reference: green cardigan | Unlocked. A beach commonly pulls the outfit toward summer clothing. | "wearing the same green cardigan" (assumed; change if she should wear beach clothing) |
| 5 | A2 | (not stated) apparent age; reference: adult | Unlocked. Age commonly drifts younger or older. | "adult woman, same apparent age" |
| 6 | A2 | (not stated) face, skin tone, eye colour; reference gives none | Unlocked and unseen. These commonly shift in derivation. | "same face, skin tone and eye colour as the reference" |
| 7 | A2 | (not stated) lighting, framing, art style, aspect ratio; reference gives none | Unlocked and unseen. These commonly change with a new place. | "keep the same framing, art style and aspect ratio" |
| — | A1, A3, A4, A5 | none found | — | — |

Note: the cafe-to-beach background change and the added "smiling" are intended changes, not drift.

### Rewritten prompt

```
Same woman as the reference, unchanged: adult, shoulder-length chestnut hair, round gold glasses, green cardigan, same face, skin tone and eye colour. Only the background and her expression change: she is now at a sunny beach, smiling. Keep the same framing, art style and aspect ratio.
```

Changed: replaced "Same woman" with explicit locks for hair, glasses, outfit, age, face and framing; stated that only background and expression change.

## P3 — mouse-on-plant stamp, mixed style

### Issues

| # | Check | Quoted words | Why it may go wrong ("commonly renders as") | Fix |
|---|---|---|---|---|
| 1 | A1 | "stamp" | Postage stamp, rubber stamp or foot action; a rubber stamp or its ink mark is a common reading. | "a postage stamp with a perforated edge" (assumed) |
| 2 | A1 | "mouse" | Animal or computer mouse; "on a plant" adds little context. | "a small field mouse (the animal)" |
| 3 | A1 | "plant" | Vegetation or factory; a factory is a plausible picture. | "a leafy green houseplant" (assumed) |
| 4 | A3 | "minimalist but extremely detailed" | Direct conflict; generators commonly pick one or blend them into noise. | Pick one: "minimalist, fine detail only in the mouse's fur" (assumed). |
| 5 | A3 | "photoreal and flat vector" | Direct conflict (photographic versus flat 2D). | Pick "flat vector" (assumed). |
| 6 | A3 | "night and day" | Two opposite lighting states in one image commonly give a muddled sky. | Generate two versions, P3a day and P3b night. |
| 7 | A3 | "stamp of a mouse on a plant, in the style of a famous living illustrator, minimalist but extremely detailed, photoreal and flat vector, night and day" | Overloaded: about 10 independent attributes (three objects, style reference, four style terms, two lighting terms). | Split as above and drop the conflicting terms. |
| 8 | A5 | "in the style of a famous living illustrator" | Points at a living artist's style; commonly reworded or refused by safety systems. The prompt contains no name, so none is flagged. | Describe the look generically: "bold outlines, limited colour palette, flat shapes" (assumed). |
| — | A4 | none found | — | — |

A2 not applicable: no reference.

### Rewritten prompt

```
P3a: A postage stamp with a perforated edge showing a small field mouse (the animal) sitting on a leafy green houseplant, minimalist flat vector style with bold outlines and a limited colour palette, fine detail only in the mouse's fur, daytime (assumed, change if wrong).
P3b: same as P3a, but under a night sky with a crescent moon.
```

Changed: resolved stamp, mouse and plant; chose minimalist flat vector over the conflicting terms; replaced the living-artist reference with a generic look; split night and day into two generations.

## Attribute-lock checklist

not applicable

## Before you spend credits

- Test P2 alone as one image first; check hair, glasses and cardigan before relying on the lock phrase.
- Confirm the assumptions marked "(assumed)" in P1 (animal bat, bird crane, six-fingered hand) and P3 (postage stamp, houseplant, flat vector) before generating.
- For P1, plan a check pass on the hand and add the "OPEN" text in an editor rather than generating it.

## Limits of this check

- Mode: P1 T, P2 R, P3 T. No reference was seen for P2, only the description given; face, skin tone, eye colour, lighting, framing, art style and aspect ratio were not described. No reference exists for P1 or P3.
- Every finding is a tendency ("commonly renders as"), not a guarantee. Nothing about any model's real behavior was verified, no images were generated, and the rewrites are untested, so try one image first.

Use cases

• Checking a prompt before a batch run so credits are not spent on a misread word
• Keeping a character identical across 10 or 20 image-to-image variants
• Finding why a generator keeps drawing the wrong "glasses", "crane" or "mouse"
• Rewording an innocent prompt that keeps getting blocked
• Spotting contradictions and overloaded prompts before you pay for retries

FAQ

Does it tell me what my model will do? No. It has not run your prompt and never claims what a specific model will do; it only says what a word commonly renders as. A one-image test render is still the real check.

Does it call an image generator or any API? No. It is read-only, makes no API call and needs no key or hosting. It keeps nothing after the conversation.

Can it see my reference image? Only if you attach it or describe it. Otherwise it works in text-only mode, lists the attributes a reference typically has, and asks you to describe yours so the drift check can be re-run.

Is it a way around safety filters? No. It only rewords innocent prompts that sound risky. If the intent itself is disallowed, it says it cannot help with that prompt.

How is this different from a shot-list planner or an image QA checklist? A shot-list planner decides which pictures a set needs, and a QA checklist judges images you already have. This one audits the prompt text before any image exists.