// Use for complex or high-risk implementation work that needs plan-first execution, explicit verification design, pre-code review, and autonomous post-implementation hardening through self-test, simplify, and review-loop.
| name | plan-loop |
| description | Use for complex or high-risk implementation work that needs plan-first execution, explicit verification design, pre-code review, and autonomous post-implementation hardening through self-test, simplify, and review-loop. |
Draft a plan, harden it with review coverage appropriate to the current host, then execute after user approval.
Read ~/dev/agent-guards/AGENTS.md before starting, plus any repo-local AGENTS.md / CLAUDE.md instructions that apply. Your plan and implementation must follow those guidelines.
plans/<task>/plan.md (for example plans/auth-refactor/plan.md) informed by research findings. Gitignore plans/. Default to clean reimplementation over patching around existing complexity — agents implement fast, so the cost of rewriting cleanly is almost always lower than the cost of maintaining a patch on bad code.self-test skill. The plan's self-test section must name the exact proof path the agent will use: commands, test files, browser/Electron automation, Computer Use steps, or a combination. For new user-facing features, prefer a real-surface smoke of the completed workflow, then use fast command checks for repeatable confidence. "Manual E2E" or "the user should verify" is not acceptable when the agent can prove the behavior itself.references/review-protocol.md. Spin up all review agents in parallel.self-test skill. Run the planned proof path on the highest-signal affected surface. If it fails, fix and re-run.simplify skill, coding agents tend to overcomplicate code.review-loop. It handles its own fix/re-verify cycle internally.Steps 8-11 run autonomously — do not pause for user input between them.
After each round, assess confidence 0-100 with top unknowns and what must be true to proceed:
<70: another round.70-84: user must accept listed risks to proceed.>=85: eligible for sign-off.self-test coverage unless the written plan names the exact proof path, environment, and blockers.self-test. plan-loop is not complete until simplify and review-loop have also run.Prove the requested goal works on the highest-signal affected surface before closing. Default closing behavior for every implementation task — not just when explicitly asked. Also triggers on: 'make sure it works', 'test it yourself', 'don't just tell me it's done', 'I don't want to test this'.
Post-implementation review/fix loop. Use host-appropriate review coverage on your implementation, fix issues, iterate until highly confident, then commit.
Lightweight read-only audit. Use host-appropriate review coverage for evidence and findings. Iterate on disagreement or low confidence. No code changes.
Iteratively clarify requirements for a new task until reaching 90%+ confidence, then auto-invoke plan-loop.
Orchestrate complex tasks with agent teams. Clarify requirements, plan with dual-model review, implement with TeamCreate, simplify, review, and ship.
Audit agent skills in a repo against best practices — checks .claude/skills/ and .agents/skills/ for quality, structure, and coverage gaps. Use when the user wants to review, improve, or evaluate their skills, or when they mention skill quality, skill audit, or skill best practices.
