Create Skill

Create an org-owned skill. Owned skills live in .claude/skills/<name>/ like framework skills, but they belong to the org: /update never overwrites them, and if upstream ever ships a skill with the same name, the org's version wins and the collision is reported.

Arguments: an optional skill name and/or a description of what the skill should do.

When to invoke

User says: "create a skill", "make a skill for X", "I want a /foo command for our team", "turn this workflow into a skill", /create-skill <name> Not this: installing someone else's published skill → find-skills · changing a framework skill's behavior → /contribute (upstream owns it)

Step 1: Name and scope

Derive a kebab-case name from the request. Rules:

• Lowercase letters, digits, hyphens only. No slashes, no leading dot.
• Must not collide with a framework skill. Check both the local tree and upstream when available:

NAME="<kebab-name>"
if [ -d ".claude/skills/$NAME" ]; then
  echo "exists locally"
fi
git fetch upstream main --quiet 2>/dev/null || true
git ls-tree -d upstream/main ".claude/skills/$NAME" 2>/dev/null | grep -q . && echo "exists upstream"

If the name exists locally or upstream, propose 2–3 alternatives via AskUserQuestion instead of silently picking one. An org CAN deliberately own a name upstream also uses (their version then always wins), but that is an explicit choice — never a default.

If the user gave only a name and no behavior, ask what the skill should do before scaffolding — a skill with an empty body helps nobody.

Step 2: Scaffold

Write .claude/skills/$NAME/SKILL.md:

---
name: <name>
description: <one line — what it does and when an agent should reach for it>
---

<What this skill does, in 1–3 sentences.>

## When to invoke

User says: <trigger phrases>
Not this: <adjacent intents that route elsewhere>

## Steps

1. <concrete step — commands in fenced bash blocks>
2. <...>

Fill it with real content derived from the user's description — trigger phrases an agent can match, concrete steps, actual commands. Follow product-voice for any user-facing copy the skill emits.

Cross-runtime adapter (mandatory). Codex invokes skills as $<name> from .codex/skills/, and Pi consumes the same tree — one adapter serves both. Write .codex/skills/$NAME/SKILL.md:

---
name: <name>
description: '<same one line>'
---

<!-- org-owned skill adapter — created by /create-skill -->

# <name> (org skill)

This org-owned skill is defined once in `.claude/skills/<name>/SKILL.md`.

1. Read `.claude/skills/<name>/SKILL.md` for the full workflow.
2. Run any referenced `bin/` scripts directly from this shell.
3. Translate interactive choices to this runtime's question tooling when
   available; otherwise render compact numbered choices with an `Other:`
   option and wait for the user.

The runtime installer only removes files it installed itself (hash-tracked), so this adapter survives Codex/Pi runtime upgrades. Without it, the skill exists for Claude Code teammates only — always write both files.

Step 3: Register ownership

Add the name to owned_skills in egregore.json (create the array if absent):

jq --arg n "$NAME" '.owned_skills = ((.owned_skills // []) + [$n] | unique)' \
  egregore.json > egregore.json.tmp && mv egregore.json.tmp egregore.json

This is what makes the skill org-owned: /update runs bin/restore-owned-skills.sh after every framework sync, which keeps every listed skill at the org's committed version.

Step 4: Share it

Tell the user the skill is ready to try now (/<name> in Claude Code, $<name> in Codex, /<name> in Pi), then run /save so it reaches the team — teammates get it on /pull or their next session start. The skill directory, the .codex/skills/ adapter, and the egregore.json change must all land in the same PR, or the ownership guarantee ships half-finished.

Telemetry (fire-and-forget):

bash bin/telemetry.sh emit "command" '{"command":"create-skill"}' 2>/dev/null &

Edge cases