Learn Skill System

A cross-platform /learn command that turns any input — code, papers, articles, sessions — into reusable agent skills with entrepreneurial-level quality judgment.

What It Does

Run /learn <source> in any supported agent and it will:

1. Parse the source (file, URL, session, pasted text)
2. Gather content via existing tools
3. Extract a reusable pattern
4. Score it using the F.L.E.C.D. rubric — a 5-dimension entrepreneurial scoring system
5. Decide whether to INVEST, DEFER, MERGE, or REJECT
6. Store in a universal intermediate representation
7. Sync to platform-native skill format
8. Report the decision with full scoring breakdown

F.L.E.C.D. Scoring

Dimension	Weight	What It Measures
Frequency	0.20	How often this pattern recurs in your work
Leverage	0.25	How many agent turns it saves per use
Edge	0.25	How unique/non-reproducible the knowledge is
Compounding	0.15	How connected it is to your existing skills
Decay	0.15	How stable the knowledge is over time

Decision thresholds:

• ≥ 8.0 → INVEST — create a full skill
• 5.0–7.9 → DEFER — store as lightweight memory, promote after 3+ matches
• 3.0–4.9 → MERGE — append to most similar existing skill
• < 3.0 → REJECT — not worth learning

Supported Sources

Source	Example
File path	/learn ~/projects/api/src/auth
URL	/learn https://arxiv.org/abs/2401.12345
Session	/learn 刚才的操作 or /learn what I just did
Pasted text	Paste content, then /learn
Requirements	/learn https://... focus on implementation

Architecture

┌─────────────────────────────────────────────────────────┐
│  /learn SKILL.md — drives the agent through 8 steps     │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  Universal Store (~/.agent-skills/)                     │
│  ├── skills/<name>/skill.json  ← source of truth (IR)   │
│  ├── ledger.jsonl              ← usage events           │
│  ├── deferred.jsonl            ← DEFERred skills        │
│  ├── decisions.log             ← audit trail            │
│  └── config.json               ← thresholds & weights   │
│                                                         │
│  Platform Adapters                                      │
│  ├── Claude Code  → ~/.claude/skills/<name>/SKILL.md   │
│  ├── Codex        → (post-MVP)                          │
│  ├── OpenCode     → (post-MVP)                          │
│  ├── QwenCode     → (post-MVP)                          │
│  └── Gemini CLI   → (post-MVP)                          │
│                                                         │
│  Stop Hook (session_end.sh)                             │
│  └── AUTO-PRUNE / AUTO-PROMOTE every 24h               │
│                                                         │
└─────────────────────────────────────────────────────────┘

Portfolio Lifecycle

Event	Trigger	Action
AUTO-PRUNE	90 days unused	Move to .deprecated/
AUTO-PROMOTE	3+ matches in deferred	Promote to full skill
AUTO-DEMOTE	60 days declining usage	Reduce priority

This Repo Contains

• openspec/ — Full OpenSpec change with proposal, design, specs, and tasks
• .claude/skills/openspec-*/ — OpenSpec commands for Claude Code
• .codex/skills/openspec-*/ — OpenSpec commands for Codex
• .gemini/skills/openspec-*/ — OpenSpec commands for Gemini CLI
• .opencode/skills/openspec-*/ — OpenSpec commands for OpenCode
• .pi/skills/openspec-*/ — OpenSpec commands for Pi
• .qwen/skills/openspec-*/ — OpenSpec commands for QwenCode

The actual /learn skill lives at ~/.claude/skills/learn/SKILL.md (installed at runtime).

Design Docs

See openspec/changes/learn-skill-system/ for:

• proposal.md — Strategic vision
• design.md — Technical architecture
• tasks.md — Implementation tasks
• specs/ — 5 capability specs

Inspiration

Inspired by Hermes Agent's /learn — extended with entrepreneurial scoring, universal IR, cross-platform sync, and automatic portfolio lifecycle management.