Reusable skills for your AI agents
365Skill is our public lab for agent skills: a standard SKILL.md format, a deny-by-default publish policy and an evals harness. Open source, installable, forkable — the exact setup we run our own agents on.
What are Agent Skills?
An agent skill is a job description for an AI agent: a directory with a SKILL.md that says when to engage, which steps to follow and which resources to load.
A standard SKILL.md format
A Markdown file with frontmatter: name and description drive triggering, the body spells out steps, constraints and output format. Executable by agents, readable by humans.
Loaded on demand
The agent sees only skill names and descriptions until a task matches — then it reads the full file. Skills can bundle instructions, deterministic scripts and references that load only when needed.
Portable across clients
An open directory format supported by mainstream coding agents — Claude Code, Codex, Cursor and others. Write once, version in git, share anywhere.
Three skills in the repo
All Experimental: structurally complete and validated, with trigger rates and field performance still being tuned.
discover-project-skills
Inventories a project's existing skills, discovers reusable practices and — with your approval — extracts them into portable skill packages. Ships with a Python structural scanner; discovery and audit are read-only.
five-step-dev
A five-step dev workflow: Research → Plan → Implement → Review → Verify. Risk-graded, plan-first, evidence-based verification — top-tier models reserved for high-stakes decisions (zh-CN).
five-step-retro
A self-iteration loop for five-step-dev: run logs plus a six-dimension retrospective produce evidence-backed proposals — only human-approved upgrades flow back into the repo (zh-CN).
Engineering discipline
Deny-by-default publishing
The public repo is a deterministic export of a private source repo: only skill IDs whitelisted in catalog/publication-policy.json ship, and CI rejects unlisted skill directories, evals and references.
Trigger evals
Every skill carries 10 positive and 10 negative cases in English and Chinese — fire when it should, stay quiet when it shouldn't. Structure checks, unit tests and real-repo scans gate every release.
Safety boundaries
The scanner never reads or outputs .env* files, credentials, private keys or secret stores, and skips dependency folders and build artifacts. Discovery and audit are read-only; writes happen only after your explicit approval.
Install in three minutes
Skills are plain directories. No package-manager magic.
Clone the repo
git clone https://github.com/laojin1900/365Skill.gitApache-2.0 — fork it, change it, republish it.
Link into your agent's skills directory
ln -s "$PWD/skills/discover-project-skills" ~/.codex/skills/Codex shown; copying the directory works too. For other clients (Claude Code, Cursor, …), drop the skill into the path their docs specify.
Invoke it
Start a task and say: use $discover-project-skills to scan this repository. See the repo README for client-specific adapters.
Useful? Give it a star
Stars, forks and issues all welcome — or see where 365Skill sits in the product family.