Author a new skill or agent for a plugin

The frontmatter spec, the loop shape, and a triggering-effectiveness review — in that order, because a skill that never fires reports no error and only a dedicated review catches it.

Also usessuperpowersclaude-plugins-official

A new skill or agent that never fires reports no error at all — it just sits unused, indistinguishable from a skill nobody needed. plugin-dev:skill-development supplies the frontmatter spec and superpowers:writing-skills the authoring loop, but neither one checks whether the description will actually trigger; that is plugin-dev:skill-reviewer's job specifically. confab:confab-agentic-reliability catches what a style/triggering review reads past — excessive tool grants and missing verify-wiring — and plugin-dev:plugin-validator runs last, once there is a real manifest to validate against.

Start here

One prompt to start

I'm writing a new skill for this plugin -- walk me through the frontmatter and structure conventions first, use the writing-skills loop to actually draft and verify it, then review the description specifically for whether it will actually trigger, audit it for agentic-reliability defects, and only validate the manifest once the components genuinely exist.

Needsplugin-dev, superpowers, confab

Beats

Run these in order. Each prompt is copy-pasteable straight into Claude Code.

1. plugin-dev:skill-developmentalso in 1 other recipe

   • Tune an instruction a workflow will reuse

   The canonical frontmatter and progressive-disclosure spec, with good-vs-bad description: examples shown side by side — the reference this whole task is judged against.

   Run this beat on its own

   I'm writing a new skill for this plugin — walk me through the frontmatter and structure conventions before I draft anything

2. superpowers:writing-skills

   Explicitly for creating new skills, editing existing ones, or verifying a skill works before deployment — the loop shape plugin-dev's reference material doesn't itself supply.

3. plugin-dev:skill-reviewer

   Reviews triggering effectiveness specifically — a skill with a weak description never fires, and nothing else in this sequence would catch that.

   Run this beat on its own

   review this skill's description and frontmatter for triggering effectiveness — will this actually fire when it should?

4. confab:confab-agentic-reliabilityalso in 1 other recipe

   • Audit your own agent and skill definitions for loop-reliability defects

   Checks for excessive tool grants and missing verify-wiring — the werkstoff-side beat that catches what a style/triggering reviewer reads past.

   Run this beat on its own

   audit this new skill/agent for agentic reliability defects — unbounded retries, missing verification wiring, tool grants beyond its stated role

5. plugin-dev:plugin-validator

   Manifest and structure, checked last, once the components actually exist — validating a manifest before the skill exists checks nothing real.

Worked example

Grounded in — why this beat order is trustworthy

This repo's own CLAUDE.md finding: frontmatter that fails to parse still loads, with no description and no tools, so the skill never triggers and nothing reports an error — which is exactly why test/plugins/lint-frontmatter.py exists and why the skill-reviewer beat here can't be skipped.

Do / Don't

• ✓Learn the canonical frontmatter and progressive-disclosure spec before drafting anything new.
• ✓Use the writing-skills loop to create, edit, and verify the skill -- the authoring loop the frontmatter spec alone doesn't supply.
• ✓Review the description specifically for triggering effectiveness -- a weak description never fires and nothing else in the sequence catches that.
• ✓Audit the new skill or agent for excessive tool grants and missing verify-wiring before shipping it.
• ✓Validate the manifest last, once the components actually exist -- validating it earlier checks nothing real.

• ×Don't skip the triggering-effectiveness review -- a skill that never fires reports no error at all, indistinguishable from a skill nobody needed.
• ×Don't validate the plugin manifest before the skill/agent files actually exist -- there's nothing real to check yet.
• ×Don't assume frontmatter that looks fine actually parses -- malformed frontmatter still loads, silently, with no description and no tools.
• ×Don't skip the agentic-reliability audit just because the description and structure already passed review -- it catches a different category of defect entirely.