AI Engineering Workflow Skills

This file mirrors AGENTS.md for compatibility.

Start Here

• Follow system-level/core.md for invariant operating rules.
• For non-trivial work, follow AI_ENGINEERING_WORKFLOW.md, including risk level, definition of done, and failure-path guidance.

Trigger Map

• Use roast before committing to build a new idea, to pressure-test it from multiple angles first.
• Use clarify-work for high-level task clarification.
• Use to-prd when scope, terminology, or success criteria are still fuzzy.
• Use decompose-to-issues before implementation on high-level work.
• Use test-driven-development (from addyosmani/agent-skills) when building features or fixing bugs where expected behavior is clear.
• Use code-simplification (from addyosmani/agent-skills) after implementing a feature.
• Use diagnose when something is broken, throwing, or regressing.
• Use security-and-hardening (from addyosmani/agent-skills) before PRs touching auth, payments, secrets, or external APIs.
• Use ai-agent-pr-metadata when configuring PR templates, PR update comments, or AI review comments to disclose agent/model/run metadata outside commit messages.
• Use open-code-review-setup when creating a new repository or onboarding an existing repo to automated AI code review (OCR), or when OCR reviews are not running on PRs.
• Use external-pr-viability before implementing a fix for any unsolicited PR against a repo we don't own — check the repo actually merges outside contributors before investing the effort.
• Use external-pr-style before opening any PR against a repo we don't own.
• Use external-campaign-triage when tracking state across many candidate issues/PRs/repos in a multi-repo external contribution campaign, and always before posting a status-check comment on a stalled external PR.
• Use resolving-merge-conflicts when you need to resolve an in-progress git merge/rebase conflict.
• Use wizard when a manual procedure needs a human to click through a dashboard or enter credentials — generates a guided script instead of leaving hitl-blocker's bare issue description to figure out alone.
• Use git-guardrails-claude-code to set up or audit the PreToolUse hook that blocks force-push, direct push to main/master/staging, reset --hard, clean -f/-fd, branch -D, and bare checkout ./restore ..
• Use docker-db-guardrails to set up or audit the PreToolUse hook that blocks schema-wipe commands (migrate:fresh, migrate:refresh, migrate:reset, db:wipe, db:drop, db:reset, DROP DATABASE/TABLE/SCHEMA, TRUNCATE, dropdb, FLUSHALL) against a Docker container unless the actually-resolved target database visibly indicates it is disposable — especially before running tests by copying files into a container.
• Use git-identity-guard to install a $PATH-level git wrapper that hard-blocks any commit or push under the wrong author/committer identity, everywhere on the machine, with no override — for when commits keep landing under the wrong email despite a documented rule.
• Use docker-volume-guardrails to set up or audit the PreToolUse hook that unconditionally blocks Docker commands that delete a volume (docker volume rm/prune, docker rm -v, docker system prune --volumes, docker compose down -v) — no override, matching the "never delete volumes without explicit human confirmation" rule.
• Use design-composition-check before a second feature builds against a freshly researched, extracted, or ingested design system that no prior screen has exercised yet — render its hardest compositions first to falsify whether it actually holds up.
• Use write-prompt-guide when a skill pack needs a user-facing PROMPT_GUIDE.md telling people what to type to get a good run out of it.
• Use writing-for-agents when creating or editing skills, or authoring agent instructions.
• Use caveman only when the user explicitly asks for caveman mode or ultra-compressed replies; keep it active until the user says "stop caveman" or "normal mode", except for Auto-Clarity safety cases.
• Use handover when context crosses an agent or session boundary, when only 5-10% of the session limit remains with work unfinished, or when context usage passes 40% on unfinished multi-step work.
• Use tmux-orphaned-socket when tmux ls/attach fails with a missing-socket error, especially right after any /tmp cleanup — check for a live orphaned server before assuming sessions are gone.
• Use vps-cleanup when a VPS or server disk is high or full (No space left on device, before a large build) — survey read-only, clear regenerable caches, then ask once, batched, before deleting /tmp worktrees/caches or Docker images; never volumes.

Summary

• Follow system-level/core.md for core engineering defaults, execution discipline, safety rules, validation, test database safety, and review traceability.
• Follow AI_ENGINEERING_WORKFLOW.md for full planning tracks, issue-scoped execution, verification gates, risk levels, and failure paths.