Image prompt refactoring
- Extract the actual goal. Separate generation from editing. Identify the subject, physical scene, and supplied references. If a critical target or conflicting requirement is unclear, ask one focused question rather than invent identity, pose, or camera metadata. Check: the goal is explicit.
- Remove or concretize noise. This workflow prefers physical constraints to vague quality inflation. Remove unsupported award/platform/superlative claims; translate meaningful style or lighting requests into observable properties. Resolution, HDR, and renderer names are not universally ineffective: they may matter to a specific model or requested workflow. Keep real output settings separate from the descriptive prompt, and preserve an explicitly requested rendering style or ask before discarding it. Check: no meaningful user constraint was silently deleted and no universal model-effect claim is made.
- Separate change from preservation. For an edit, select one bounded edit objective. Preserve applicable identity, facial structure, pose, garment, framing, perspective, and lighting except properties explicitly changed. Do not assert unknown exact focal lengths. If compositing needs coordinated lighting changes, name those within the same objective, never as invariants. For a new image use a creation target, not fictional original-image invariants. Check: nothing is simultaneously changed and preserved.
- Assign references by function. Give each image one distinct extraction role, such as subject identity/garment or environment geometry/materials. Explicitly exclude unintended transfer. Read references/worked-examples.md when the roles or edit boundary need illustration. Check: every supplied reference has a role; conflicting roles are resolved or queried.
- Return this inspectable schema. Omit reference roles when none exist and omit output settings when none were requested. For generation, preservation can state only user-specified invariants or say no original image exists.
CORE INTENT: <Physical summary of goal>
[TARGET TO CHANGE]: <One bounded edit objective, or creation target>
[PRESERVATION LIST]: <Applicable unchanged properties; exclude changes>
[REFERENCE ROLES]:
- Reference Image 1: Extract strictly <distinct role; exclusions if needed>.
[OUTPUT SETTINGS]: <Requested supported settings, separate from scene description>
Done: the prompt makes the target, invariants, and reference roles explicit; no contradictions or invented metadata remain. This skill produces text only, not images. Prompt structure is not a guarantee of deterministic model obedience.