← ClaudeAtlas

self-update-skillslisted

Audit and safely improve any existing agent skill when authoritative documentation, product features, APIs, policies, standards, user corrections, observed failures, or better workflows make it stale or incomplete. Use when asked to refresh, modernize, maintain, self-evaluate, or self-update a skill; when new information discovered during another task may affect a relevant skill; or when checking a skill library for drift. Research and propose changes first, require explicit user approval before editing, then make narrow updates with validation, regression checks, provenance, and rollback. This skill governs updates to other skills as well as itself.
apnatvar/web-development-and-online-marketing-skills · ★ 0 · AI & Automation · score 60
Install: claude install-skill apnatvar/web-development-and-online-marketing-skills
# Self-Update Skills Maintain skills as versioned, evidence-backed operational artifacts. Discovering an improvement is permission to investigate and propose—not permission to edit. ## Non-negotiable contract - Apply this workflow to any relevant skill, not only this skill. - Never modify, delete, move, publish, install, or commit a target skill before presenting a concrete proposal and receiving explicit user approval for that proposal. - Treat approval as scoped to the listed target, files, changes, tests, and external actions. Re-propose material scope changes. - Keep the current version recoverable through version control or an exact backup. Never destroy the last known-good version. - Prefer authoritative, current evidence over novelty, popularity, or an agent's unsupported judgment. - Preserve user-authored rules and repository instructions unless the user explicitly approves changing them. - Do not weaken safety, compliance, approval, privacy, or validation requirements merely to make a workflow faster. - Do not claim that a revision is better without comparative evidence. ## Controlled update workflow ### 1. Identify a real trigger Accept update candidates from: - an official feature, API, platform, policy, standard, or dependency change; - a deprecation, security advisory, compatibility break, or changed provider requirement; - a user correction or explicit maintenance request; - a reproducible failure, inefficient trace, missing edge case, or repeated workaro