auto-claude-skills

Use when executing implementation plans with 3+ independent, file-disjoint tasks that benefit from parallel specialist agents with shared contracts and reviewer-gated completion

Use when verifying an implementation still matches its spec or plan — during REVIEW or SHIP, or on demand to check drift, confirm you are still on plan, or run a spec check — surfacing spec deviations, unvalidated assumptions, and untested code paths against Intent Truth

Use when mining the repo for improvement proposals in the LEARN phase — manually sweeping eval baselines, gate-status output, memory feedback, and parked revival criteria into a ranked, evidence-graded proposal report with in-session approve/reject and a GitHub-issue queue

Use when you need to run the repo's own declared test/lint/type gate locally and emit pass/fail evidence — during REVIEW, before requesting code review, or on a request to run the tests or verify the build — discovering the gate from CLAUDE.md, Makefile, pyproject, or .verify.yml

Use when you need to prove a change actually works through its real interfaces — during REVIEW or on requests like validate the feature, does it work, run e2e, or smoke test — covering browser E2E, API smoke, CLI checks, and a11y, perf (Lighthouse), and visual-regression audits with graceful tool-degradation

Use when starting a new feature or initiative and you need problem context, prior art, and acceptance criteria before design — the DISCOVER phase entry point — pulling Jira/Confluence context and synthesizing a discovery brief to validate problem framing

Use when a design or implementation involves autonomous agents, unattended/background operation, private-data access combined with external/untrusted input, or outbound actions (sending data, posting, pushing, API calls) — the lethal-trifecta risk

Use when authoring or de-generic-ifying persuasive prose — essays, blog posts, op-eds, newsletters, talks, or a "make this not sound like AI" rewrite — to apply a post-draft authorial-judgment revision pass. NOT for README/API-docs, specs/changelogs, or code.

Use when a code change touches 5+ files or modifies auth/secrets/permissions/hooks/CI paths and needs multi-lens parallel review (security, quality, spec, governance) before merge.

Use when any SDLC phase needs external docs, dependency internals, cross-session memory, or feature specs — provides tiered context retrieval across External Truth (docs), Internal Truth (dependencies), Historical Truth (memory), and Intent Truth (feature specs) with graceful degradation based on installed tools.

Use when preparing to ship or release — before pushing to production, promoting a build, or finalizing a branch — to confirm CI is green, no WIP commits remain, and version and design artifacts are in order

Use when investigating production symptoms — connection failures, pod crashes/restarts, SIGTERM/OOM errors, latency spikes, Cloud SQL/proxy issues, deployment-correlated errors, ImagePullBackOff, CreateContainerConfigError, or node NotReady events

Use when creating new skills, commands, or plugins — emits repo-native seed files (SKILL.md skeleton, routing entry, test snippets)

Use when transforming, migrating, refactoring, or generating across many files at once — codebase-wide renames, 50+ file migrations, mass test/doc generation, framework upgrades — via claude -p with manifest, dry-run, and log-based retry

Use when reviewing a shipped feature's real-world outcome in the LEARN phase — checking adoption, error, or experiment metrics after release, validating ship-time hypotheses, or deciding follow-up work — querying PostHog and creating gated follow-up Jira work

Use when shipping a completed feature and generating as-built OpenSpec docs before branch finalization

Use when reviewing code changes for security issues — during REVIEW phase or on explicit security, vulnerability, SAST, or secret-scan requests — running a STRIDE threat-model pre-pass, then available Semgrep/Opengrep, Trivy, and Gitleaks scanners with a self-healing fix loop

Use when capturing a durable, team-relevant learning into the committed .claude/knowledge/ base — a gotcha, decision, convention, or runbook worth sharing with teammates' agents. Human-gated.

Use when investigating a published supply-chain attack on a registry package (npm, Maven, PyPI, Go, Gradle) — advisory-driven org-wide audit. Triggers on attack-language ("compromised", "malicious", "hijacked", "backdoored", "typosquatted"). NOT for routine CVE scanning — that routes to security-scanner.

Use when facing flapping alerts, alert fatigue, recurring noisy incidents, threshold audits, or SLO-alert redesign questions for a GCP monitoring project.

Multi-Agent Debate (MAD) for complex designs. Spawns architect, critic, and pragmatist for collaborative design exploration with structured convergence.

Produce 3 thin comparable variants of a proposed design with a comparison artifact and mandatory Human Validation Plan

On-demand postmortem trend analysis — recurrence grouping, trigger categorization, MTTR/MTTD from canonical docs/postmortems/ corpus