prompt-qc
Quality-check generation prompts for self-containment, canon, model-specific format, reference labels, exact text and dialogue, timing, camera contradictions, continuity, and prompt slop.
SKILL.mdskills/prompt-qc/SKILL.md
GitHubPrompt QC
Workflow
- Read core contract, style rules, dramaturgy rules when the source is visual, continuity rules, reference asset rules, prompt packet schema, the source brief, and the target adapter skill.
- Check each prompt for:
- source ID present
- target model and mode correct
- every subject and reference identified with the correct role
- no "same as before" dependency
- exact visible text and dialogue preserved
- duration and timeline agree
- camera instructions are physically compatible and motivated by the shot brief
- lighting does not contradict itself
- character, wardrobe, prop, and location continuity match canon
- no canonical
must_not_betrait is requested - required speech/movement/stillness intent is not accidentally contradicted
- required visible-dialogue line, speaker, mouth visibility, cut policy, and timing survive adaptation
- required end-frame state and capture behavior survive without conflicting camera instructions
- negative constraints do not fight positive instructions
- model-specific fields and section order are correct
- for MiniMax H3,
h3_base_skillish3-prompt-writing, the applicable style skill was selected (or explicitly recorded asnone), the style overlay did not replace the H3 field structure, the final prompt follows the selected base/Ref2VA guide, andscripts/minimax_h3_prompt_validator.pypasses for the actual mode and duration before generation
- Fix prompt-level defects directly. Escalate canon conflicts to the upstream artifact.
- Run the style checker over prompt files.
Done
Every prompt can be handed alone to a generation operator with its named reference media and still produce the intended task definition.