Learn from your workflow patterns and suggest skills, hooks, and agents to automate repetitive work.
Claude Code:
/plugin marketplace add agent-sh/agentsys
/plugin install skillers@agentsysOther tools (OpenCode, Codex CLI, Cursor, Kiro): npm install -g agentsys && agentsys install skillers.
- Work normally - your AI tool saves conversation transcripts automatically (Claude Code, Codex, OpenCode)
/skillers compact- Analyze transcripts and extract recurring patterns into weighted knowledge/skillers recommend- Get actionable suggestions for automation
| Command | Description |
|---|---|
/skillers show |
Show status, transcript stats, and knowledge themes |
/skillers compact [--days=N] |
Analyze transcripts and extract patterns (20 sessions per run; run again for more) |
/skillers recommend |
Suggest skills, hooks, and agents to create |
--days defaults to 7. Every command takes --scope=global|repo|both (or --global, --repo); the default is global.
- Pain points ("this auth flow is broken again")
- Repeated requests ("run tests after editing auth/")
- Task themes ("refactoring the login module")
- Wishes ("I wish the AI knew our API patterns")
- Workflow sequences (multi-step patterns you follow consistently)
- API keys, passwords, or secrets
- One-off tasks that won't recur
- Normal productive work without friction
scripts/skillers.js does the deterministic work: it reads Claude Code, Codex and OpenCode transcripts, redacts secrets, writes a sampled digest, and later validates, weights, merges and prunes the knowledge files and applies the evidence bar. The model does the judgment: the compactor agent reads the digest and writes observations, the recommender agent picks what is worth automating. Nothing is created without your pick.
Knowledge lives in ~/.claude/skillers/ (global) or <repo>/.claude/skillers/ (repo scope, which only learns from sessions run inside that repo). $AI_STATE_DIR replaces .claude. Requires Node.js; reading OpenCode needs Node 22.5 or newer.