new-skill

Scaffold one new skill, fast and correctly. The goal is a valid, dogfood-ready skill folder on disk in a single interaction - not a full eval loop. If the user wants rigorous, benchmark-driven iteration, point them at the skill-creator skill.

The skill name may be passed as an argument ($ARGUMENTS). If absent, ask for it.

1. Capture intent

Ask only what you cannot infer from the conversation (if the user just said "turn this into a skill", mine the history first). Keep it to a short interview:

1. One job - what should this skill enable, in one sentence? If it is two jobs, split into two skills.
2. Triggers - what would a user actually say or be doing when this should fire? Collect real phrasings.
3. Output - what should the skill produce or do (format, steps, side effects)?
4. Support files - does it need a scripts/ helper, references/ docs, or assets/? Usually no.

2. Settle the name

The folder name equals the name field equals the slash command. It must be:

• kebab-case: lowercase a-z, 0-9, hyphens; no leading, trailing, or consecutive hyphens; 64 chars max
• free of the reserved words anthropic and claude
• distinct from existing skills (check siblings and installed skills)

Confirm the name with the user before writing anything.

3. Locate the repo root

The new skill should land beside the others, where the helper scripts live. This skill is usually installed in ~/.claude/skills/ but run from inside the repo, so find the repo from the working directory - NOT from this skill's own location:

• Search the current working directory and its ancestors for the nearest directory containing CONVENTIONS.md, _template/, and scripts/validate.sh. That is REPO_ROOT. New skills go as siblings: REPO_ROOT/NAME/.
• If none is found, treat it as standalone use: create ./NAME/ in the current directory (or ask where the skill should live) and use the manual checklist in step 5 instead of the repo scripts.

4. Write the SKILL.md

Create TARGET/NAME/SKILL.md. Prefer copying REPO_ROOT/_template/SKILL.md and filling it in. If the template is not reachable (standalone use), write this skeleton (replace each brace placeholder; keep it plain markdown, no angle-bracket tags):

---
name: {folder-name}
description: {what it does}. Use when {natural trigger phrases the user would say}.
---

# {Title}

{One line: the single job.}

## Instructions

{Imperative steps. State what to do, not why - the body stays in context across
turns once invoked, so every line is recurring cost.}

1. ...
2. ...
3. {output or what to return}

Crafting the description (this is the trigger and the most important field):

• Lead with the key use case, then what it does, then when - woven with the real trigger phrases from step 1.
• Lean slightly pushy - skills under-trigger. e.g. append "Use this whenever the user mentions X, Y, or Z, even if they don't explicitly ask for it."
• Keep it 1024 characters max (hard cap in the portable standard).
• Use no angle-bracket or XML-tag content in the name or description - it is disallowed by claude.ai and the API. Use square brackets or quotes for placeholders.
• Avoid a literal : (colon then space) in an unquoted description - it can break YAML parsing. Rephrase, or quote the whole value.
• Put "when to use" only in the description.

Frontmatter keys (this is what most often breaks loading): include ONLY keys Claude Code recognizes - name, description, and optionally argument-hint, when_to_use, arguments, allowed-tools, disallowed-tools, disable-model-invocation, user-invocable, model, effort, context, agent, hooks, paths, shell. Do NOT add license, metadata, or compatibility: Claude Code silently hides a skill from the / menu if it sees an unrecognized key. The repo's license lives in the top-level LICENSE file, not in frontmatter.

Crafting the body: imperative, concise, under 500 lines. Briefly explain why a step matters rather than stacking rigid all-caps rules - the model follows reasoning better than rules. Move anything long into references/ and link it.

If support files are needed, create the scripts/, references/, or assets/ dirs and reference them from SKILL.md using the ${CLAUDE_SKILL_DIR} path prefix so paths resolve wherever the skill is installed.

5. Lint it

After the SKILL.md is written, run the repo linter with the Bash tool. Do NOT use a shell-injection block for this - injection runs when this skill loads, before the new skill exists, so it would lint nothing. If REPO_ROOT was found in step 3, run:

bash REPO_ROOT/scripts/validate.sh NAME

If the validator is not reachable (standalone use), verify by hand against this checklist:

• name matches the folder; valid kebab-case; 64 chars max; no reserved words
• frontmatter has only Claude Code-recognized keys (no license, metadata, compatibility)
• description non-empty, 1024 chars max, leads with use case plus trigger phrases, no XML tags, no unquoted :
• body under 500 lines
• no secrets or keys committed
• one clear job

Fix anything that fails and re-check.

6. Install and verify it loads

Passing the linter proves the file is well-formed; it does NOT prove Claude Code will surface it. Close that loop:

• If REPO_ROOT/scripts/install.sh exists, install with it (it re-lints, then copies). Add --symlink to keep the repo as the single source of truth:

bash REPO_ROOT/scripts/install.sh NAME

Otherwise copy the folder into the personal skills dir: cp -r NAME ~/.claude/skills/NAME.

• Tell the user to FULLY quit and relaunch Claude Code. A brand-new skill appears in the / menu only after a restart; edits to an already-loaded skill hot-reload without one.
• Have them confirm it shows: type /NAME, or run /doctor to list discovered skills. If it does not appear, the usual cause is an unrecognized frontmatter key - re-check step 4.

7. Wrap up

• If an IDEAS.md exists at the repo root, offer to update this skill's row (set its status to building, or add a new row if it was not listed).
• If a README.md with a skills table exists at the repo root, offer to add a row for the new skill (name, one-line description, install command).
• Remind the user to dogfood it in real work before publishing - that is the cheapest test.
• For sharing, others install it the same way: copy the folder into ~/.claude/skills/ (or use scripts/install.sh), then restart.
• If they want trigger-accuracy tuning or benchmark-grade evals, hand off to the skill-creator skill - this skill intentionally stops at a solid, linted draft.