スキル一覧に戻る
richardhowes

reflection

by richardhowes

0🍴 0📅 2026年1月11日
GitHubで見るManusで実行

SKILL.md


name: reflection description: Self-improving meta-skill that learns from user corrections and updates the TARGET PROJECT's skill files. Only runs when user explicitly says "reflect". Commits all skill updates to Git.

Reflection Skill

Purpose

Improve future Claude Code sessions by learning from mistakes and user corrections. Updates skill files in the target project's .claude/skills/ directory.

This is a meta-skill - it modifies other skills based on session learnings.

When to Run (Explicit Triggers Only)

Only activate when the user explicitly requests reflection:

  • "reflect"
  • "don't do this again"
  • "remember this for next time"
  • "update your skills"
  • "learn from this"

Never run automatically - always wait for explicit user trigger.

Where Rules Go

Rules are written to the target project's skill files (NOT the plugin repository):

Rule TypeTarget File
General conventions.claude/skills/conventions.md
VILT patterns.claude/skills/vilt-patterns.md
Project-specific.claude/skills/project-rules.md
Unclear/mixed.claude/skills/lessons-learned.md

If the target project doesn't have .claude/skills/, create it with a starter structure.

Process

  1. Analyse the completed session

    • Review conversation for corrections
    • Identify repeated mistakes
    • Note stated preferences
  2. Extract atomic, durable rules

    • One rule per concept
    • "Do X" or "Never Y" format
    • Date-stamped for tracking
  3. Check for contradictions

    • Read existing skill files
    • Flag conflicting rules
    • Ask user to resolve
  4. Update appropriate skill file(s)

    • Append to ## Learned Rules section
    • Preserve existing formatting
    • Keep changes minimal
  5. Commit changes separately from code

    • Use semantic commit message
    • Include rule summary in body

Rule Quality Bar

A rule must be:

  • Durable: Will apply to future work in this project
  • Actionable: Clear imperative ("Always X", "Never Y")
  • Specific: Not too broad or vague
  • Non-contradicting: Doesn't conflict with existing rules

Rule Format

- [YYYY-MM-DD] RULE_TEXT (source: BRIEF_CORRECTION_CONTEXT)

Examples:

- [2025-01-09] Always eager load relationships in service methods that return collections (source: booking service review)
- [2025-01-09] Never use axios - use Inertia router instead (source: corrected 3x in dashboard component)
- [2025-01-10] Use FloatingLabelInput for all form fields (source: user preference stated)

Starter Skill Structure

When creating .claude/skills/ in a new project:

# Lessons Learned

## Purpose
Project-specific rules learned from Claude Code sessions.

## Learned Rules
<!-- Reflection appends here -->

Git Commit Format

chore(skills): reflect session learnings

- Added: [rule summary]
- Source: [correction context]
- File: .claude/skills/[filename].md

What NOT to Learn

Do not create rules for:

  • One-time exceptions
  • User's momentary preferences
  • Context-specific decisions
  • Temporary workarounds

Only learn patterns that should always apply.

スコア

総合スコア

45/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

0/10
説明文

100文字以上の説明がある

0/10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

0/5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です