Agent Skills
› doodledood/manifest-dev
› auto-optimize-prompt
auto-optimize-prompt
GitHub通过循环调用审查与工程技能，迭代优化提示词直至无问题。支持原子输出、用户交互解决歧义及收敛报告。
Trigger Scenarios
需要优化或修复提示词
提示词存在模糊或错误需迭代修正
Install
npx skills add doodledood/manifest-dev --skill auto-optimize-prompt -g -y
SKILL.md
Frontmatter
{
"name": "auto-optimize-prompt",
"metadata": {
"internal": true
},
"description": "Iteratively auto-optimize a prompt until no issues remain. Uses review-prompt in a loop, asks user for ambiguities, applies fixes via prompt-engineering skill. Runs until converged."
}
Auto-Optimize Prompt
User request: $ARGUMENTS
Iteratively optimize a prompt until no issues remain.
Goal
Loop until a general-purpose agent that activates the review-prompt skill finds no issues: review → resolve NEEDS_USER_INPUT with user → fix via prompt-engineering → repeat.
- No path provided: Ask which file to optimize
- Working copy: Use
/tmp/auto-optimize-*.mdduring iterations; apply to original only when converged
Constraints
| Constraint | Why |
|---|---|
| Converge, don't cap | No iteration limits—run until no issues |
| Atomic output | Original unchanged until fully converged |
| DRY | Delegate review to a general-purpose agent activating the review-prompt skill, fixes to prompt-engineering |
| User-in-the-loop | NEEDS_USER_INPUT issues require user resolution (with context, options); skip if user declines |
Output
Report: file path, iterations, issues fixed (auto vs user-resolved), issues skipped, summary of changes.
Version History
- 2ac487e Current 2026-08-20 08:55
