What does General Image Prompt Adapter do?
Positioning
Input: blueprint (from character-design-engine / illustration-direction) + target model name (optional).
Output: a natural-language prompt in English (code block), and—only when the target officially supports it—parameter lines outside the prompt.
Does not: design. If the input is not a blueprint (no proposition with a verb, outline, four-layer clothing, a puncture point, locked facts), it falls back to aesthetic-director-core.
Does not: invent model behavior. The models covered by this Skill are numerous and change quickly; if uncertain, write [Unverified], do not guess parameters.
Hard Rules
- Target model unknown and affects format → ask one question: "Which model will render the image? If unsure, write a generic natural-language prompt." Do not list options.
- The prompt body contains only visual descriptions.
--ar, weights, samplers, negative prompts, etc., do not go into the body unless the target's official documentation supports them and the owner requests it. - Any claim that "model X is more sensitive to Y" must carry an evidence label. Every entry in
references/target-notes.mdis already annotated; targets not in the table are always[Unverified]. - Do not add generic quality words like
masterpiece, 8k, ultra-detailed, award-winning. - Do not modify locked facts. Anything removed by the blueprint must not reappear in any form.
Execution Flow
Step 1 Confirm target owner specifies / read from Session State / ask one question / default to generic
Step 2 Check target notes references/target-notes.md: format preferences, length, parameter support, review evidence labels per entry
Step 3 Organize order Subject and framing → Proposition structure (outline spatial relations) → Appearance → Four layers of clothing (shape+position+behavior)
→ Pose causality and instant → Environment (keep only the layer that explains light/action) → Main light source and attenuation → Color palette three levels and puncture point location
Step 4 Write prompt Coherent English paragraph (not comma-separated tags, unless the target notes say tags are more stable).
Prioritize relational words: over / beneath / descending from / wrapping / weighed down by / the only saturated color is
Step 5 Parameter lines Only if the target officially supports them: on a separate line, outside the code block or at the end of the code block on its own line, annotated with [Official]
Step 6 Length generic target: 60–120 words. If too long, first remove repeated adjectives, minor accessories, background details; do not remove proposition, anchor points, puncture points, causality, locked facts
Step 7 Self-check
Output Contract
Direction: <blueprint proposition, one line>
Target: <model name or generic>
```text
<English natural-language prompt>
Parameters (if applicable): <e.g. --ar 3:4 --no jewelry> [Official] / N/A
Loss and Remediation: <what relation this format loses, how it compensates, one to two lines>
In Standard mode, same as `anima-prompt-compiler`: V1 faithful + V2 enhanced, V2 does not change V1 facts.
## Trade-offs When Translating the Blueprint
Natural language format retains relationships best, but the main risks are **model rewriting** (some services rewrite prompts before generation) and **length decay** (tail information is ignored):
- Place locked facts and puncture points within the first 40 words.
- Each color word is placed immediately next to its noun and appears only once.
- "The remaining oddity" is written as a complete short sentence, not a single word—single words get rewritten away.
- Do not write "no X" (many models treat negations as positive); instead, use positive constraints: `bare neck and wrists` instead of `no jewelry`. Use parameters only when the target officially supports exclusion parameters (e.g., MJ `--no`).
## Self-Check
- [ ] Input is a blueprint?
- [ ] Target confirmed or explicitly marked generic?
- [ ] Body has no parameters, no quality words, no negative sentences?
- [ ] Locked facts and puncture points within the first 40 words?
- [ ] Every model assertion has an evidence label?
- [ ] Deleted items have not returned?
- [ ] Has a "Loss and Remediation" line?
## References
- `references/target-notes.md` — Format notes for common targets, each annotated with evidence labels; targets outside the table are always [Unverified]