meta

atrislabs 的头像作者 atrislabs

人工智能代理的元认知技能。在开始工作、感觉卡住、输出感觉不佳或执行复杂任务之前使用。教导如何思考思考。

📖 中文导读:人工智能代理的元认知技能。在开始工作、感觉卡住、输出感觉不佳或执行复杂任务之前使用。教导如何思考思考。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: meta description: Metacognition skill for AI agents. Use when starting work, feeling stuck, output feels off, or before complex tasks. Teaches how to think about thinking.

Meta Skill

How to be an effective agent. Load this when you need to check yourself.

When to activate

  • Starting a new session (orientation)
  • Feeling stuck or uncertain
  • Output feels sloppy or generic
  • Complex task ahead
  • After a mistake

Context awareness

  1. MAP first - Read atris/MAP.md before any search. It's the index.
  2. Journal context - Check today's log for recent work, patterns, blockers.
  3. LESSONS.md - What failed before? Don't repeat.
  4. Skill routing - Frontend? Load design. Backend? Load backend. Writing? Load writing.

Pace control

Signal Action
Simple task, clear path Execute fast
Complex task, multiple approaches Slow down, plan first
Uncertain about intent Ask ONE question
Output feels off Pause, re-read context
Stuck > 2 attempts Escalate, don't spiral

Self-correction triggers

Stop and reassess when:

  • You're about to search without checking MAP
  • You're writing code during PLAN phase
  • You're adding features not requested
  • You're explaining instead of doing
  • You've used words like "robust", "comprehensive", "streamlined"

Memory leverage

Before acting:

  1. Check atris/policies/LESSONS.md - short lessons from past mistakes
  2. Check atris/features/ - prior art for similar work
  3. Check journal history - grep -i "<keyword>" atris/logs/**/*.md

Skill composition

You can stack skills:

  • meta + design = thoughtful frontend work
  • meta + backend = careful API design
  • meta + writing = deliberate prose

Meta is the foundation. Domain skills are the specialization.

The loop

Orient (read context)
    ↓
Decide (which skill? what pace?)
    ↓
Act (execute with focus)
    ↓
Check (output match intent?)
    ↓
Learn (add to LESSONS if miss)

Anti-patterns

  • Grepping before MAP check
  • Coding during planning
  • Adding unrequested features
  • Verbose explanations
  • Guessing instead of asking
  • Repeating past mistakes
  • Creating skill without symlinking to .claude/skills/ (won't load mid-session)

Skill evolution

Skills improve from your mistakes:

  1. After REVIEW, log learnings to LESSONS.md (format: date | skill | lesson)
  2. Repeated lessons (2-3x) get promoted into the relevant skill
  3. Next agent loads the improved skill

You're training future agents. Leave the codebase smarter than you found it.

One-liner

Think before you act. Check yourself. Use the system.