Create Skill

Create Skill

Create one new Muse skill. Use this skill only for explicit Muse skill creation requests, not for ordinary skill usage, code changes, benchmark tasks, or third-party skill/plugin systems.

Two scopes exist (ADR 8975):

• Project scope (the default): the skill lives in the current workspace at .agents/skills/<skill-id>/.
• Personal scope (the user asked for a "personal", "user", or "cross-project" skill): the skill belongs in the managed personal root $CONFIG_DIR/skills/<skill-id> ($XDG_CONFIG_HOME/muse, else $HOME/.config/muse). You stage and validate the draft in the workspace, then hand the user one muse skills install command — the store performs the managed install (files + provenance), so skills update and skills uninstall keep working on it.

Never write to a foreign harness root (~/.codex/skills, ~/.claude/skills) or to $HOME/.agents/skills — those are import-only sources, never write targets. Never write directly into $CONFIG_DIR/skills either: the store owns that write, through the install command below.

Scope

• Create exactly one directory: .agents/skills/<skill-id>/ (project scope) or the staging directory .agents/skill-drafts/<skill-id>/ (personal scope — deliberately OUTSIDE .agents/skills/, so the draft is not loaded as a project skill).
• Create exactly one required file: <that directory>/SKILL.md.
• Do not create a plugin, install a plugin, enable a skill, trust a plugin, execute the generated skill, fetch remote content, or write outside the current workspace.
• Do not add scripts, assets, references, or extra files unless the user explicitly asks for them and the target remains inside the new skill directory.

Inputs

Before writing files, identify:

• scope: project (default) or personal — personal only when the user asked for a personal/user/cross-project skill.
• skill-id: a portable lowercase identifier for the directory and frontmatter name.
• description: one clear sentence for the frontmatter.
• body: concise instructions that make the generated skill useful on its own.

Ask a short clarification question if the user did not provide enough information to choose a safe skill-id and useful behavior.

Safety Checks

Reject the request before writing when:

• the destination is not under .agents/skills/ (project scope) or .agents/skill-drafts/ (personal staging) in the current workspace;
• the requested final destination is a foreign harness root (~/.codex/skills, ~/.claude/skills), $HOME/.agents/skills, or any absolute path outside the two sanctioned roots and the personal staging destination (the workspace .agents/skills/ tree, $CONFIG_DIR/skills/<skill-id>, and .agents/skill-drafts/<skill-id>) — explain the sanctioned path instead;
• the ID is empty, ., .., contains / or \, starts with -, or contains anything except ASCII lowercase letters, digits, hyphen, or underscore;
• the ID is a Windows reserved stem such as con, prn, aux, nul, com1, com2, com3, com4, com5, com6, com7, com8, com9, lpt1, lpt2, lpt3, lpt4, lpt5, lpt6, lpt7, lpt8, or lpt9;
• the destination already exists, is a symlink, or any parent resolves outside the current workspace.

Creation Steps

Use normal file and shell tools with the current workspace as the base. <dir> below is .agents/skills/<skill-id> (project scope) or .agents/skill-drafts/<skill-id> (personal scope).

1. Check that <dir> does not exist.

2. Create its parent (.agents/skills or .agents/skill-drafts) if needed.

3. Reserve the leaf directory with a no-replace operation. If another process wins the race, stop and report incomplete.

4. Recheck that the reserved directory resolves inside the current workspace and is not a symlink.

5. Write <dir>/SKILL.md.

6. Run:

   muse skills validate <dir> --json
   

7. Treat the draft as complete only when the validator succeeds, returns valid: true, and reports zero diagnostics or warnings.

8. If validation fails, correct the same draft and revalidate. Stop after three correction rounds and report the remaining validator output as incomplete.

Generated SKILL.md

Use this shape:

---
name: <skill-id>
description: <one sentence>
---

# <Readable Skill Name>

<Instructions for when and how to use the skill.>

Keep the generated instructions direct and self-contained. Include only behavior the user asked for or that is necessary for the skill to work.

Completion Report

On success, report:

• the canonical path to the created directory;

• the validator command and clean result;

• that no install, enable, trust, or execution step was run;

• personal scope only: the one command that finishes the managed install into the personal root — run by the user, so the skills store records the install provenance itself:

  muse skills install .agents/skill-drafts/<skill-id>
  

  and that the staging directory can be deleted after the install succeeds.

On failure, report:

• what operation failed;
• the destination if it was reserved;
• the remaining diagnostics or tool error;
• which checks were not completed.