cjl-autoresearch-cc

Overview

Improve skills, prompts, articles, workflows, and systems via iterative mutation-testing.

Core principle: One small verifiable change per round. Large rewrites are unverifiable and will be reverted.

Workflow: small edits → test → score → keep improvements, discard regressions.

Inspired by Karpathy/autoresearch and 0xcjl/openclaw-autoresearch-pro.

Trigger Keywords

English: autoresearch

Chinese: 自动优化, 自动研究

Semantic Triggers (No Keywords Needed)

This skill activates when the user's intent matches, even without explicit keywords:

• User wants to improve any skill, prompt, article, workflow, or system
• User asks to polish, refine, enhance, or upgrade content
• User wants iterative testing and improvement
• User says '帮我改进一下这个prompt', 'optimize this'
• User says '迭代优化', '循环改进', '反复打磨'
• User asks '能不能更好', '如何提升质量'
• User uses "打磨", "精炼", "完善", "升级" in context of content improvement

Supported Optimization Targets

Workflow

Step 1 — Identify Mode and Target

Before proceeding, confirm with user:

"Optimize [target] in [mode] mode? (yes/no)"

If no, ask for clarification. If yes, proceed to Step 2.

Parse the user's request to determine mode. Check for:

Keyword triggers:

• autoresearch [target] / 自动优化 [target] / 自动研究 [target]
• optimize [target] / improve [target] / 优化 [target] / 改进 [target]
• refine [target] / enhance [target] / 精炼 [target] / 增强 [target]

Semantic triggers (intent-based):

• User wants to improve any skill, prompt, article, workflow, or system
• User asks to polish, refine, enhance, or upgrade content
• User describes wanting iterative testing and improvement

Mode detection from intent:

For Skill/Plugin mode, resolve the path:

• Skill: ~/.claude/skills/\x3Cskill-name>/SKILL.md
• Plugin: ~/.claude/plugins/\x3Cplugin-name>/

If path doesn't exist, search in order: ~/.claude/skills/ → current dir → ask user.

Examples of semantic triggers (no keywords):

• "帮我优化一下这个skill" → Skill mode
• "这个prompt不够好，帮我改进" → Prompt mode
• "我想让这篇文章更通顺" → Article mode
• "优化一下部署流程" → Workflow mode

Step 2 — Generate Checklist (10 Questions)

Read the target content first. Then generate 10 diverse, specific yes/no checklist questions relevant to the content type:

For Skill/Plugin mode:

For Prompt mode:

For Article/Documentation mode:

For Workflow/System mode:

Present the 10 questions, numbered 1-10. Ask the user to select which ones to activate.

Rule: Must use at least 5 questions. Using fewer makes scoring unreliable.

After presenting, ask: "Ready to start the optimization loop? (yes/start)"

Step 3 — Prepare Test Cases

Test cases validate that mutations improve, not harm, the content. Generate realistic user scenarios.

• Skill/Plugin mode: Generate 3-5 realistic prompts a user would send when using the skill/plugin
• Prompt mode: Generate 3-5 test inputs that the prompt would process
• Article mode: Generate 3-5 ways the article might be read or consumed
• Workflow mode: Generate 3-5 scenarios the workflow would handle
• System mode: Generate 3-5 conditions the system would encounter

Store test cases in context — do not write to disk unless needed.

Step 4 — Run Autoresearch Loop

Tip: For mutation strategies, see Mutation Strategy Reference below.

Loop configuration:

• Rounds per batch: 30
• Max total rounds: 100
• Pause: After every 30 rounds, show summary and ask user to continue or stop
• Stop conditions:
  • User says stop
  • 100 rounds completed
  • Score reaches 100%
  • No improvement for 10 consecutive rounds

Per-round procedure:

Track progress: Round N/100 | Best: XX% | Last: +/-YY

Constraint: ONE mutation per round. Multiple changes = unverifiable = will be reverted.

1. Mutate: Make ONE small edit (see Mutation types)

2. Test: For each test case, simulate what output the content would produce

   Constraint: Be honest. If the output would disappoint a user, the mutation failed.

3. Score: Apply each active checklist question (0 or 1 per question). Score = (passed / total_questions) × 100

   Scoring scale:

   • 10/10 = 100% (perfect)
   • 7/10 = 70% (good)
   • 5/10 = 50% (minimum viable)

4. Decide: If new score ≥ best score → keep the mutation. If lower → revert

   Example: Best=85%, New=87% → Keep. Best=85%, New=83% → Revert.

   Trust the score. Don't rationalize a bad mutation.

5. Log: Round number, mutation type, score, keep/revert decision

Mutation types (pick ONE per round):

Step 5 — Report Results (after each batch)

See Quick Reference below for output format examples.

After each batch (30 rounds):

Example:

Batch 1 (rounds 1-30):
  Best score: 85%
  Mutations kept: 23  |  Reverted: 7
  Most effective types: A, C, D

After full completion:

Optimized: [filename/path]
Score: XX% → YY% (+ZZ%)
Rounds: N (kept: K, reverted: R)
Top mutations: [type, type, type]
---
Final content:
[diff or inline]

Mutation Strategy Reference

High-impact, low-risk changes:

• Adding explicit constraints where the content is vague
• Expanding coverage to cover edge cases
• Adding concrete examples to abstract instructions
• Tightening soft language ("try to" → "must")

Avoid in one round:

• Large rewrites of entire sections
• Multiple unrelated changes at once
• Changing fundamental scope or purpose
• Formatting-only changes (no testable value)
• Adding content the user didn't request
• Removing more than 10% of content

Quick Reference

Keywords Reference

Auto-detect: autoresearch, 自动优化, 自动研究 Skill: autoresearch ~/.claude/skills/tdd Prompt: optimize this prompt: [text] Workflow: optimize the deployment workflow System: improve the error handling system

Semantic Triggers (No Keywords)

"帮我优化一下这个skill"           # → Skill mode
"这个prompt不太行"               # → Prompt mode
"我想让文章更通顺"               # → Article mode
"优化一下部署流程"               # → Workflow mode
"改进一下这个系统"               # → System mode
"improve this code review"         # → Prompt/Skill mode
"polish this documentation"       # → Article mode

Mode Detection

Edge cases: Empty → ask. Invalid path → fallback to ~/.claude/skills/. Ambiguous → ask.