lalachan-story-improvement-loop

Use when a LALACHAN story needs repeated critique and revision cycles, especially 10-round polish, shareability repair, natural Chinese dialogue, Xiaoyunque story readiness, or running a Codex exec lo

SKILL.md

GitHub

LALACHAN Story Improvement Loop

Use this skill when a story still feels strange, hard to understand, too AI-like, not shareable, or not ready for video generation.

This skill extends the lalachan-story-critic rules. It is for repeated improvement, not a single rewrite.

Core Loop

Run the story through explicit rounds:

draft -> exact problem list -> revised story -> next focus

Default to 10 rounds for important stories. Use fewer rounds only for tiny repairs.

Round Checklist

Use one main focus per round:

  1. Story promise: what the audience expects to see.
  2. Duration fit: 15s needs one joke; 30s can support setup, twist, payoff.
  3. Causality: every event should be caused by an action or visible rule.
  4. Character voice: each line must sound speakable by that character.
  5. Dialogue read-aloud: remove translated, slogan-like, report-like lines.
  6. Visual comedy: one memorable image people can retell.
  7. Safety and credibility: dangerous actions need cartoon framing or setup.
  8. Prop function: every prop gets one job or stays background.
  9. Share hook: end with payoff or a clear next question.
  10. Final compression: remove repeated explanations and prompt leakage.

Failure Modes

Reject:

  • abstract morals that characters would not say
  • events connected only by 突然
  • too many props for the duration
  • fake epic battle language for small jokes
  • system-log robot dialogue
  • prompt instructions inside the story text
  • unclear physical actions that may render as unsafe

Output Shape

For a documented loop:

## Round N
Problem:
Change:
Result:

For a final deliverable:

## Final Story
...

## Prompt Notes
...

Keep Prompt Notes separate from the story.

Codex Exec Script

In the LALACHAN repo, use:

scripts/lalachan_story_improve_loop_codex.sh STORY.md --rounds 10

The script uses:

model: gpt-5.5
reasoning effort: xhigh

It saves every prompt and result under:

references/story-improvement-runs/

Use --dry-run to inspect prompts without spending model credits.