← ClaudeAtlas

skill-craftlisted

Use for creating, revising, or read-only auditing an Agent Skill: SKILL.md, triggers, progressive disclosure, fixtures, overlap, and token cost. Defaults to audit; edits need explicit authority. Not for plugin packaging or delivery; use implement.
mblauberg/provenant · ★ 2 · AI & Automation · score 68
Install: claude install-skill mblauberg/provenant
# Skill craft A skill is a reusable, triggerable procedure, not a solved-task narrative and not a place to park static reference. Every skill earns its slot on four pillars: **Trigger** (fires on the right requests, stays silent on adjacent ones), **Structure** (frontmatter contract, progressive disclosure, soft length budgets), **Steering** (leading words and bright-line rules over repeated emphasis), and **Pruning** (cutting sediment, duplication, sprawl and no-ops before they ship). This skill has two branches and defaults to the safer one. - New or materially revised skill → [references/author.md](references/author.md). - Assessment of an existing skill, read-only → [references/audit.md](references/audit.md). - **Default with no explicit authoring request: audit.** Do not cross audit → edit without an authority envelope naming `implement` as action-owner; surface that requirement instead of silently fixing what an audit finds. ## Shared doctrine Both branches follow this without restating it in two voices: - **Three-mode invocation:** direct audit; direct authoring under authority; or composed, where audit is a companion to a primary lifecycle owner (for example `implement`) that stays the action-owner. - **Budgets:** soft ~500-word body per branch file; the standing catalogue-description lever is governed in `MAINTAINING.md`; link it, don't restate the number here. - **Token model, three-tier:** always-loaded frontmatter, triggered branch body, one-