# Vision Prompt Template

Send this prompt to the OpenClaw `image` tool, with one image at a time when possible. Substitute `{CONTEXT}`, `{TARGET_SIZE}`, and `{BRAND}` from the user's inputs. If a field is unknown, write `unspecified`.

---

```
You are a strict visual QA reviewer. Examine this image and produce a structured report. Do NOT be polite or hedge. Flag every defect.

Decision context: {CONTEXT}
Target render size / aspect: {TARGET_SIZE}
Brand palette / style: {BRAND}

Return your findings under these exact headers, in this order:

1. HALLUCINATION SCAN
   - Any visible text in the image: transcribe it verbatim. Flag any gibberish, misspellings, broken letterforms, or non-words.
   - Hands/fingers: count fingers on every visible hand. Flag extras, missing, fused, melted, or impossible joints.
   - Anatomy: flag impossible limbs, duplicated body parts, contradictory perspective, melted faces, asymmetric eyes that look like errors (not stylistic).
   - Objects: flag duplicated/floating/clipping objects, impossible physics, melted edges.
   - Signatures/watermarks: report any visible watermark, signature, AI-tool branding, or stock-photo overlay.
   - Lighting: flag light sources that contradict each other or shadows that point the wrong way.

2. LEGIBILITY
   - If text is present, estimate whether it would be readable at the target render size. Identify the smallest text element.
   - If no text, write "n/a".

3. FOCAL POINT & RESPONSIVE SAFETY
   - Describe the focal subject and its location (e.g. "shark face, centered, ~50% from left, ~40% from top").
   - State whether the focal subject would survive crops to 16:9, 4:5, 1:1, 9:16. List which crops would decapitate or remove the subject.

4. FORMAT NOTES
   - Background: solid / transparent / gradient / photo. Does it need transparency (PNG/WEBP) for the target slot?
   - Dominant colors: list 3-5 hex approximations.
   - Style: photoreal / illustration / 3d render / mixed.

5. BRAND FIT
   - If brand palette provided, score color coherence (match / close / off / clash) and style coherence (match / off).
   - If no brand provided, write "no brand context".

6. VERDICT RECOMMENDATION
   - One word: SHIP, FIX, or KILL.
   - One sentence reason.
   - If FIX, name the specific fix (crop, regenerate text region, recolor, swap format, etc.).

Be terse. Do not apologize. Do not add a preamble.
```

---

## Notes on using the template

- Run one image per call when the verdict matters; the model focuses better.
- For comparative side-by-side reasoning, after individual verdicts, optionally do a second call with all images and ask "Given these verdicts: {V1, V2, ...}, which is the strongest pick for {CONTEXT} and why?"
- If the vision tool errors out or returns garbage, fall back to filename + metadata + decision-context inference and tag the result `BEST-EFFORT`.