outcomefocusai

claude-learn

The self-improving plugin for Claude Code. Claude gets measurably better every session — automatically. Your proven learnings help every user. Every user's learnings help you.

---

What This Does

Claude Code starts every session from zero. You correct the same mistakes. It retries the same dead ends. Good approaches vanish. claude-learn fixes this.

It captures behavioral rules from real outcomes, scores them with evidence, validates across sessions, and loads them automatically. Rules that work persist. Flukes decay. Clusters of rules graduate into skills. And your proven rules feed a community playbook that makes everyone's Claude better.

Observe → Capture → Score → Validate → Decay or Confirm → Prune or Graduate
    ↑         ↑         ↑                                        |
    |     [3 layers]  [evidence]                                  |
    └─────────────────────────────────────────────────────────────┘

Your proven rules ──→ Community Playbook ──→ All users benefit

---

Collective Intelligence

This is the headline feature no other tool has.

When your rules reach "proven" status (validated across 3+ sessions, score 3.0+), you can contribute them to the community playbook. Other users receive them on their next plugin update. Their Claude validates the rules independently — good rules survive, bad ones decay per-user.

A rule validated by many independent users across different projects is essentially a universal law of AI coding assistance.

How It Works — The Full Loop

You use Claude normally
    ↓
Rules accumulate in your personal playbook (scored, validated)
    ↓
Rules reach "proven" status (score 3.0+, confirmed across 3+ sessions)
    ↓
Run /learn contribute
    ↓
Claude does TWO things:
    ├── 1. Proposes your new proven rules for community
    └── 2. Reports your validation scores on EXISTING community rules
            ("read-before-edit is score 4.2 on my system, confirmed 6x")
    ↓
Submitted as GitHub issue (or PR if you have write access)
    ↓
Maintainer reviews:
    ├── Merges new quality rules at score 1.0
    └── Bumps existing rules when 3+ users independently validate them
    ↓
Plugin update delivers changes to ALL users
    ↓
Community rule scores reflect real validation:
    1 contributor  → score 1.0 (unproven, must earn trust on your system)
    3+ validators  → score 2.0 (independently confirmed by multiple users)
    5+ validators  → score 3.0 (widely proven — high confidence)

The key insight: Every time someone runs /learn contribute, they're not just proposing new rules — they're validating existing ones. This is what makes community scores go UP. More users validating = higher scores = more trust for new installs.

How to Contribute

# In Claude Code, run:
/learn contribute

# Claude will:
# 1. Show your proven rules (score 3.0+, 3+ sessions) → proposes new ones
# 2. Check your scores on existing community rules → reports validations
# 3. Strip personal details, generalize wording
# 4. Create a GitHub issue with new rules + validation report

# For repo contributors with write access:
# Edit templates/playbook-community.md directly and submit a PR

You cannot modify the core plugin code. Only the maintainer can merge changes. Contributors can only submit issues/PRs which require approval. Your contributions add value — they never risk breaking anything.

Two Playbooks, One System

Playbook	File	Source	Priority
Personal	~/.claude/rules/playbook.md	Your sessions	Highest — always wins conflicts
Community	~/.claude/rules/playbook-community.md	All contributors	Starting point — must prove itself to YOU

Both auto-load every session. Personal rules always take priority. Community rules start at their community-validated score (1.0 to 3.0 based on how many users confirmed them), but can still decay on YOUR system if they don't fit your workflow.

---

Why This Is Better Than Every Alternative