claude-self-improve 0.1.0

Scaffold a .claude/ governance framework for self-improving AI agents

claude-self-improve

Scaffold a .claude/ governance framework that makes AI coding agents self-improving.

What it does

claude-self-improve init creates a .claude/ directory in your git repo containing:

• 4 hook scripts that enforce governance automatically (session tracking, anti-pattern reminders, edit tracking, reflection gates)
• 7 meta-skill commands that teach the agent to learn from mistakes, stay in scope, and improve its own skills
• Memory files for persistent learning across sessions (anti-patterns, skill registry, session logs)
• CLAUDE.md template with governance rules pre-configured

How it works

Hooks fire automatically         Meta-skills guide agent cognition
────────────────────────         ─────────────────────────────────
SessionStart → remind agent      /meta-anti-patterns → avoid past mistakes
PreToolUse   → check compliance  /meta-scope-guard   → stay within request
PostToolUse  → track edits       /meta-learn         → reflect after work
Stop         → enforce learning  /meta-commit        → learn + commit atomically

                    Memory persists across sessions
                    ───────────────────────────────
                    MEMORY.md      → skill registry, insights
                    ANTI_PATTERN.md → documented mistakes
                    session logs    → longitudinal patterns

The agent improves itself: /meta-learn captures lessons, /meta-self-audit discovers patterns across sessions, and /meta-evolve proposes governance improvements — all with human approval.

Install

pip install claude-self-improve

Quick start

cd your-project
claude-self-improve init
git add .claude/
git commit -m "feat: add AI agent governance framework"

# One-time setup per developer machine:
claude-self-improve link

What gets created

.claude/
├── CLAUDE.md                  # Agent conventions (customize for your project)
├── settings.json              # Hook configuration
├── hooks/                     # 4 governance enforcement scripts
│   ├── session-init.sh        # SessionStart: init tracker + reminder
│   ├── pre-edit-governance.sh # PreToolUse: anti-pattern reminder
│   ├── post-edit-tracker.sh   # PostToolUse: track edited files
│   └── stop-reflection-gate.sh# Stop: block if no learning reflection
├── commands/                  # 7 meta-skill slash commands
│   ├── meta-anti-patterns.md  # Review documented mistakes before edits
│   ├── meta-scope-guard.md    # Define and hold scope boundaries
│   ├── meta-learn.md          # Post-task reflection loop
│   ├── meta-commit.md         # Reflect + commit atomically
│   ├── meta-propose-skill.md  # Create/improve reusable skills
│   ├── meta-self-audit.md     # Discover patterns across sessions
│   └── meta-evolve.md         # Propose governance improvements
└── memory/                    # Persistent learning across sessions
    ├── MEMORY.md              # Skill registry + project insights
    ├── ANTI_PATTERN.md        # Documented error patterns
    ├── skill-candidates.md    # Patterns awaiting promotion
    ├── topic-index.md         # Index of detailed memory files
    └── sessions/              # Per-user session logs

Customization

After init, customize these files for your project:

• .claude/CLAUDE.md — Add your project overview, directory structure, and conventions
• .claude/memory/MEMORY.md — Will auto-populate as the agent learns your project
• .claude/commands/ — Add domain-specific skills alongside the meta-skills (e.g., debug-build.md, deploy.md)

CLI reference

claude-self-improve init [--target DIR] [--force]
    Scaffold .claude/ governance framework.
    --target DIR   Target directory (default: current directory)
    --force        Overwrite existing .claude/ directory

claude-self-improve link [--target DIR]
    Create memory symlink for Claude Code's native memory system.
    --target DIR   Project directory containing .claude/ (default: current directory)

Requirements

• Python >= 3.9
• A git repository (recommended, not required)
• Claude Code or compatible AI coding agent

License

MIT