meta

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

⭐48stars🔀2forks📁View on GitHub🕐Updated Jan 7, 2026⬇️1downloads

When & Why to Use This Skill

This Claude skill implements a sophisticated metacognition framework that enables AI agents to 'think about thinking.' It optimizes agent performance by providing structured protocols for context awareness, pace control, and self-correction. By leveraging historical lessons and session journals, it prevents repetitive errors and ensures the agent remains aligned with complex task requirements, significantly improving the reliability and efficiency of autonomous workflows.

Use Cases

  • Complex Task Orchestration: Use when starting multi-step projects to ensure the agent creates a plan and checks context before executing code.
  • Error Prevention & Recovery: Activate when an agent feels 'stuck' or starts looping, providing a framework to pause, reassess intent, and escalate if necessary.
  • Continuous Improvement: Automatically reference a 'LESSONS.md' file to ensure the agent learns from past mistakes and doesn't repeat them in new sessions.
  • Contextual Navigation: Streamline information retrieval by enforcing a 'MAP-first' search strategy, reducing token waste and improving search accuracy in large codebases.
  • Output Quality Control: Trigger self-correction when responses become too verbose, generic, or drift away from the user's original instructions.
namemeta
descriptionMetacognition 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.