skillnerds/xskill/using-xskill

Using xskill

Overview

xskill distills reusable Skills (SKILL.md folders) out of the real execution trajectories of coding agents (Claude Code, Codex, OpenCode, Cursor, …). A background daemon watches each agent's session logs, slices them into single-intent atoms, clusters atoms into skills, and writes/versions each skill in its own git folder. New skill versions only replace old ones when real traffic shows they serve users better (canary A/B by UX score) — not by an LLM grading itself.

Core mental model: raw trajectory → atoms → candidate routing → SKILL.md → canary A/B → installed into every agent's skill dir. You operate the daemon; the daemon does the distilling.

When to Use

• Installing xskill or filling in ~/.xskill/config.yaml (LLM + embedding endpoints)
• Starting/keeping the daemon running (xskill serve), or backfilling old trajectories
• Joining or hosting a team server (xskill serve --server / xskill connect)
• Understanding the agent pipeline, atoms, canary/UX scoring, or deployment modes
• Re-distilling the whole skill library (e.g. after switching to a stronger model)

Quick Reference

The daemon is the engine: most commands only change state in the DB; nothing is distilled unless xskill serve (or the team server) is running.

Progressive Disclosure — read on demand

• Install & configure (config.yaml fields, per-agent collect/install paths, team

client setup): references/installation.md

• How it works (TaskAgent → TaskClusterAgent → SkillEditAgent, atoms, canary/UX

scoring, standalone vs team mode): references/mechanisms.md

• Rebuild the skill library (a ready-to-run prompt that walks a model through

re-distilling correctly): references/rebuilding-skill-library.md

Common Mistakes

• Running `rebuild` with no daemon up. rebuild only resets DB state; the watcher

in serve does the actual re-split/re-cluster every 30s. No daemon = nothing happens.

• Deleting raw `~/.xskill/*_sessions/*.md`. Those are the input to distillation —

delete them and you can no longer rebuild.

• Expecting DeepSeek to do embeddings. DeepSeek has no embedding endpoint; point the

embedding: block at DashScope / OpenAI / Ollama.

• Putting tokens in public places. Team join tokens must never land in a public repo

or chat log.