Skill Forge

Forge a new skill or agent end-to-end using the skill-creator CLI. The forge is the toolchain that scaffolds, validates, and ships skills and agents; this skill is the playbook that drives it.

When to invoke

Trigger on any of:

• "create a skill for X" / "author a new skill"
• "scaffold a skill" / "make me a skill that …"
• "add an agent" / "author a subagent for X"
• "ship this skill" / "run the skill through its gates"

Do not trigger for:

• Editing an existing skill's guidance — that is skill-creator refine <name> (corrections, not a new artifact).
• Retiring a skill — that is skill-creator skill retire <name> (suggest-first, reversible).
• Forging a cartridge (content / department / chipset bundle) — that is the cartridge-forge skill.

Prerequisites

1. docs/SKILL-FORMAT.md is the normative frontmatter/format reference; docs/skill-authoring/skill-md-quality-checks.md lists the hard defects the gate rejects; docs/authoring-tools.md maps the CLI surface. Read them once per session if the context is not already loaded.
2. Pick the slug (kebab-case, matches the skill name:). Keep it specific — the description and triggers must not collide with an existing skill's activation.
3. Decide skill vs agent. A skill is auto-loading guidance (a SKILL.md); an agent is a delegated subagent with its own tool set (an agents/<name>.md). Author the one the task needs.

The Forge Loop

create → fill → validate → critique → ship → publish

Every step is idempotent. skill ship is the gate — if it exits non-zero, the skill is not ready to publish.

1. Create

skill-creator create

The wizard scaffolds a spec-valid SKILL.md: repo-canonical frontmatter (version / format / status: ACTIVE) plus a body seeded from a template preset (workflow-guide / reference / checklist) rather than a blank prompt. It prints a Next: validate → critique → ship footer.

For an agent instead of a skill:

skill-creator agents create --name=<slug> --description="…"

(or run it interactively). It delegates to the AgentGenerator and writes a spec-valid agent .md with comma-separated tools.

2. Fill

Replace the template body with real guidance. A well-formed skill body typically wants:

• A one-line intro naming what the skill drives.
• A When to invoke section with concrete trigger phrases and an explicit do-not list (the highest-signal way to keep two on-demand skills from co-activating).
• The working steps — commands, rules, or a decision procedure.
• An Output contract if the skill produces a deliverable.

Keep the frontmatter description and triggers distinct: a trigger that is a verbatim-truncated copy of the description is a hard defect (see step 3).

3. Validate

skill-creator validate <name>

Runs the skill-content gate: it blocks on structural threats (unsafe YAML, path traversal, schema violation, secret in frontmatter) and sanitizes-and-warns on in-body issues. A clean exit is required.

4. Critique

skill-creator critique <name>

Surfaces quality issues (thin triggers, missing sections, vague description) that validate does not hard-block. Address the findings before shipping.

5. Ship

skill-creator skill ship <name>

Runs the ship pipeline in order — validate → critique → test-triggering — with a per-gate verdict. Exit code 0 means <name> is ship-ready. This is the equivalent of cartridge-forge's eval gate: do not report "done" before it is green. Use --mock to dry-run the pipeline.

For an agent, the parallel gate is:

skill-creator agents validate

which now exits non-zero on spec defects.

6. Publish

skill-creator publish <name>

Promotes the ship-ready skill. Only run after skill ship is green.

Rules

1. Never hand-edit a skill under .claude/skills/. That tree is gitignored and self-mod-guarded; the tracked source is project-claude/skills/. Author there and let node project-claude/install.cjs deploy it.
2. Never skip skill ship. It is the publish gate.
3. A correction is not a new skill. Route "the skill got X wrong" to skill-creator refine, not a fresh create.
4. Keep triggers distinct from the description and from other skills' triggers — co-activation is a real cost on every session.
5. Commit the SOURCE only (project-claude/…), never .claude/… — the git-add-blocker enforces this.

Reference walkthrough

A well-formed recent skill to model: project-claude/skills/skill-frontmatter-doctor/SKILL.md — a single-file, on-demand skill with a tight description, distinct triggers, and a clear defect-detection procedure.

Output contract

After the loop completes, report to the user:

• Final name, artifact kind (skill / agent), and source path.
• Frontmatter shape: version / format / status.
• validate / critique / ship verdicts (each shown green).
• Whether it was deployed (install.cjs) and published.

Do not report "done" before skill ship (or agents validate) has been run and shown green.