Standard: 011
Category: Practice
Status: Draft
Applies to: Every repository worked by an agent
The Focus AI
2026-07-25
Revision R1
Agent skills
1. Introduction
Skills are the capabilities an agent brings to a repository, and they are installed rather than written into it. That makes them a dependency like any other: pinned, reproducible, and restored by mise install rather than by someone remembering which ones a project needs.
This standard covers how skills are installed and kept current. Which skills belong to the standard set is an inventory that changes on its own schedule and lives in best-practices/GDE-008-skills.md, not here — a list of names is data, and putting data in a standard means revising the standard whenever the data moves.
Out of scope. The standard skill set itself. How to write a skill. What each skill does.
2. Requirements language
The key words MUST, MUST NOT, SHOULD, SHOULD NOT and MAY in this document are to be interpreted as in RFC 2119 [1]. A clause marked DEPRECATED was normative in an earlier revision and is retained so that repositories still running it can find out what replaced it.
3. Requirements
A skill MUST be installed with the skills CLI.
Never by copying a directory. A hand-copied skill has no lock entry, no hash, and no way to be restored on another machine — it works for whoever copied it and silently does not exist for anyone else.
skills add The-Focus-AI/focus-ai-brandEvery installed skill MUST be recorded in skills-lock.json, and the lock file MUST be committed.
The lock is what makes an install reproducible: it pins each skill by hash, and it is the reason a fresh clone and a six-month-old checkout resolve to the same capabilities. It is package-lock.json for agent behaviour.
mise run install MUST restore skills from the lock.
skills experimental_install reads the lock and restores by hash, which is what makes STD-004 §3.5 true for skills as well as packages. An agent environment that has to be told to install skills separately will be the one that was not.
A lock entry name MUST be kebab-case and match its skill folder.
Title-case duplicates alongside kebab-case names produce two entries that install the same skill twice and disagree about which is current.
Removing a skill MUST remove its lock entry.
skills rm deletes the files and leaves the entry, so the next experimental_install puts it back. A skill that reappears after being removed looks like a tooling fault and gets worked around rather than fixed.
A skill update MUST be reviewed as a diff before it is committed.
skills update rewrites hashes across the lock. Committing that unread accepts a behaviour change to every agent that touches the repository, sight unseen — the closest thing this fleet has to a supply-chain event.
A repository MUST NOT install a skill set wholesale from a source that carries unrelated skills.
Adding a whole upstream collection pulls in whatever else lives there. Install the named skills the standard set calls for, and nothing else.
4. References
[1] Bradner, S., Key words for use in RFCs to Indicate Requirement Levels, BCP 14, RFC 2119, March 1997.
[2] The Focus AI, STD-004 — Repository tooling and tasks. Draft.
[3] The Focus AI, best-practices/GDE-008-skills.md — the standard skill set inventory.