Create Local Skill

Create Local Skill

Invoke as $create-local-skill.

Scaffold a new user-local skill directly into ~/.codex/skills/<name>/ (and optionally ~/.claude/skills/<name>/) as a real directory. The upstream agentic-skills repo is not touched. At the end, offer to promote the skill by copying it into the user's personal fork.

npx skillpacks uninstall-global only removes skillpacks-owned skill installs whose source is inside the agentic-skills repo, so user-authored real directories are safe from upstream sync.

Process

1. Parse <skill-name> and optional description from arguments. If missing, ask for:

   • Skill name (kebab-case)
   • One-line description
   • Type (default execution)
   • Whether to also create a Claude version in ~/.claude/skills/<name>/

2. Validate $HOME/.codex/skills/<name>:

   • If it is a symlink into agentic-skills/base/ or agentic-skills/packs/, refuse — it would shadow a repo-managed skill.
   • If it is a real dir or user-owned symlink, confirm before overwriting.

3. Write ~/.codex/skills/<name>/SKILL.md:

   ---
   name: <name>
   description: <description>
   type: <type>
   version: v0.0
   ---
   
   # <Title>
   
   Invoke as `$<name>`.
   
   <1-2 sentence purpose statement.>
   
   ## Workflow
   
   1. <step>
   2. <step>
   3. <step>
   
   ## Output
   
   <what the skill produces>
   

4. Optionally mirror to ~/.claude/skills/<name>/SKILL.md (Claude frontmatter convention — no Invoke as line; add argument-hint if relevant).

5. Report created paths and remind the user of the runner-specific reload path. Claude Code should run /reload-skills first; /clear starts a new empty-context conversation and can pick up the refreshed registry; restart if the top-level .claude/skills directory did not exist at session start or the skill is still invisible. Codex should start a fresh Codex CLI session if the $ skill list remains stale.

6. Ask: "Would you like to promote this skill to your personal fork of agentic-skills?"

   If yes:

   • Ask for the path to their fork. If they don't have one, explain: fork on GitHub and clone, or git remote set-url origin <their-repo> to repoint an existing clone.
   • Verify with git -C <path> rev-parse --show-toplevel.
   • Copy the skill into <fork>/base/codex/<name>/ (and <fork>/base/claude/<name>/ if applicable), or into a <fork>/personal/... subtree if the user prefers to segregate personal skills.
   • If the skill is promoted into a current agentic-skills fork's base/ or packs/ tree, refresh that fork's public skills catalog export by running node scripts/generate-skills-catalog-export.mjs and scripts/validate-skills-catalog-export.sh.
   • If the promoted skill needs benchmark coverage, update the separate agentic-skills-benchmarks fork/repo: import the fork's export, then add either a deterministic custom setup under tests/layer4/setups/ or an explicit blocked row with blocked_reason and next_command.
   • For promoted custom setups, include a deterministic quality rubric when practical, or record why quality scoring is blocked/deferred instead of adding a subjective rubric.
   • Recommend pnpm catalog:check and pnpm bench:coverage from the benchmark repo before benchmark coverage changes are committed.
   • Run git status in the fork and suggest a commit message. Do not commit or push — leave that to the user.

7. Safety note: pushing to the upstream agentic-skills repo requires write access. Plain clones without access are rejected at the remote; forks push to the user's own repo. Cannot route a user's experimental skill into the shared upstream.

Notes

• Never write into this repo's base/ or packs/ directories. Target paths are always under $HOME/.codex/skills, $HOME/.claude/skills, or an explicitly-supplied personal fork path.
• Do not create extra docs beyond SKILL.md unless asked.

Default Shipping Contract

Follow the shared shipping contract convention in CLAUDE.md.