Plan Brief: requirement → grilled plan → launcher prompt → review (审查闭环)

One job, three artifacts: (1) a detailed execution plan document, (2) a copy-paste launcher prompt that starts a FRESH session to execute it, (3) later, the review of that session's report. You plan and review; you NEVER execute the plan and never spawn execution sub-agents.

Apply goal-first: each planned investigation or step must serve the agreed outcome or resolve a decision-critical unknown. Carry budgets, stopping conditions and how to handle incidental findings into the existing plan/prompt; do not add another process document.

Deliverables checklist — none of these may be skipped silently:

• plan doc docs/plans/YYYY-MM-DD-<slug>.md (the project's own working-doc dir if it differs)
• launcher prompt: fenced inside the plan doc AND printed in chat
• when the report comes back: <slug>.review.md beside the plan, plus a fix prompt if fixes are needed

When this skill applies

• The user wants a durable execution plan, execution handed to a separate session, or a report reviewed / a fix round dispatched — the three artifacts above.
• The user's explicit choice wins; never switch workflows silently. In-session implementation is quick-do / steady-do; if a handoff would help but was not requested, propose it before changing who executes.

Phase 0 — Absorb the requirement

• Read the raw requirement completely (chat text, or the file the user points at).
• Restate in 2-3 sentences: goal, scope, observable completion criteria, what "done" looks like.
• Split it into explicit requirements vs open details. Do NOT plan yet.

Phase 1 — Survey the project

• Read project rules (AGENTS.md), README, project memory and docs, entry points, and everything the requirement touches.
• Establish: what exists today, the real call chains, existing capabilities and reuse points, compatibility boundaries, and this project's own conventions (env managers, doc locations, run commands, who launches formal runs).
• Investigate the facts yourself from code, config, memory, docs, and upstream source/issues. Never invent a command, API detail, or platform limit; ask the user for decisions, never for facts you can inspect.

Phase 2 — Grill every open detail

Resolve only material details still missing after investigation. Use this bounded round-based method directly — it needs no separate skill:

• Ask only unresolved decisions that change scope, behavior, compatibility, risk, execution order, or acceptance; give a recommendation with brief trade-offs.
• Group independently answerable questions into one round; defer dependent ones until their prerequisites settle. After each round, recompute what it unblocked and ask the next: rounds continue while a further question would still change a decision above. The stopping rule is "nothing material left open", not "one round done".
• Do not repeat settled questions, invent a quota, or assume every requirement is incomplete.

ALWAYS settle the following, from investigation or an explicit user decision:

• Runtime environment — exact interpreter and environment (venv/conda path + activation command), env var names, working directory, package manager, versions. NEVER assume system Python. Point at existing secret locations without copying credentials into the plan.
• Inputs/outputs — formats, paths, expected artifacts, compatibility constraints.
• Execution order — dependencies, milestones, genuinely independent or parallel work.
• Document location — the project's working-doc directory; default docs/plans/YYYY-MM-DD-<slug>.md. Report and review live beside the plan with the same basename.
• Verification mode — mandatory decision; name it in one line in the plan and justify it:
  • TDD mode — when explicitly requested or required by project policy; invoke the tdd skill and identify a meaningful verification seam. A fast feedback loop alone does not require TDD.
  • Smoke-and-read mode (default for research, long-running, or heavy-compute work): the smallest real execution path + rereading your own code + checking the adjacent behavior it touches. No pytest, no test scaffolding, no wrapper harnesses — the run itself is the cost. The user runs the real workload afterwards.
  • Rule of thumb: if writing and running the test costs more than the failures it would catch, don't write it.
• Acceptance and responsibility — what is observed, the expected result, cost and environment limits, and who runs/verifies what at the end.

Recap the agreed scope, approach, and verification briefly, then write the plan without a separate approval gate — wait only if the recap introduced a new material choice.

Phase 3 — Set the taste

Apply ponytail directly (spelling the rules out here is enough):

• Plan the smallest working change: reuse existing code, then stdlib/native features, then installed dependencies. No over-encapsulation, speculative abstraction, defensive boilerplate, redundant re-validation, future scaffolding, or unrelated cleanup.
• Fix root causes along the real call chains and account for sibling callers of shared logic. Preserve trust-boundary validation, security, accessibility, data protection, and error handling that prevents data loss.
• Include authorization gates for destructive or irreversible operations; permission to execute the plan is not permission to overwrite secrets or change unrelated environments. Functional code that runs correctly is DONE.

Phase 4 — Write the plan document

One structured Markdown document in the user's language (Chinese headings by default: 目标与范围 / 现状 / 环境与前提 / 执行步骤 / 注意事项与已知坑 / 验证方式 / 交付物清单). Tables for comparisons and parameters, lists for steps, no heading followed by a wall of text. Sections:

1. Goal and scope — the Phase 0 restatement, including non-goals.
2. Current state — what exists now and which parts get reused; Phase 1 findings.
3. Environment and prerequisites — exact known commands, paths, versions; unresolved external blockers labeled as such.
4. Execution steps — ordered; each step states what / the files or command / the expected result / its dependencies. Before bulk generation or a broad matrix, settle source/version and semantic contracts, then validate a representative vertical slice and shared components. Group independent work; do not weaken the final agreed coverage. Do not invent commands or API details that were not verified.
5. Constraints and known pitfalls — compatibility, risks, and the authorization gates from Phase 3.
6. Verification — name the chosen mode (TDD / smoke-and-read) and why; concrete acceptance evidence and who runs what. Distinguish static checks, successful loading, local calls, and real user interaction; never treat absence of errors as proof of the outcome.
7. Deliverables — include <plan-basename>.report.md beside the plan. Required product docs and project memory still apply.

Phase 5 — Write the launcher prompt

Append it to the plan doc inside a fenced block AND print it in chat for copying. It must be fully self-contained for a fresh session:

• Point to the plan doc path; instruct the executor to read it plus AGENTS.md and the relevant project docs before acting.
• Carry the agreed scope, environment, step order and dependencies, ponytail taste, verification mode, authorization gates, and deliverables. Follow the plan instead of restarting workflow selection or re-grilling settled decisions.
• Keep current constraints, evidence paths and the active edit unit in the existing todo/context; recover only missing or changed material. Read exact source before editing, refresh after a match failure, and edit the generator rather than giant generated output when a source path exists.
• Hard constraints: activate the specified environment before running anything; follow the plan's step order; no scope creep; stop and ask when the plan is ambiguous, new evidence invalidates it, or scope must change; resolve ordinary local implementation details without repeated approval.
• Development report is mandatory — even for partial work: write <plan-basename>.report.md beside the plan, in the user's language, with 完成情况总表 (per plan step: 完成 / 部分 / 搁置 / 未动 + evidence), 做了什么, 没做什么与搁置原因, 规范遵循情况, 特殊处理与偏离, 遗留问题与建议.
• Partial completion is legitimate: steps blocked by real constraints (missing dependencies, environment limits, out of scope) may be shelved — never fake completion, never silently skip. The report says exactly what was and was not done.
• Evidence honesty: separate static checks, successful loading, injected input, and real end-to-end behavior. If something depends on the user's environment, give the concrete remaining check and the expected observation.
• State the deliverables.

Review loop (审查闭环) — when the user brings the report back

When the user returns saying the executor finished (or drops the report path), audit — do not re-explore the project from scratch:

1. Read the plan doc and the report side by side. Spot-check the code only where a claim needs evidence.
2. Build a 核对表: per plan step — report claim / evidence / verdict (符合 / 存疑 / 不符).
3. Audit 规范遵循: environment activation, step order, scope, ponytail taste, chosen verification mode, authorization gates, test-script restraint. Judge each deviation with its reason instead of assuming every deviation is wrong.
4. Verdict (符合 / 部分符合 / 不符合), then concrete prioritized minimal fixes in ponytail style.
5. Write <slug>.review.md beside the plan in the user's language and print it in chat. NEVER edit the executor's report — it is their artifact.
6. If fixes are needed, produce a new self-contained launcher prompt (Phase 5 rules) re-dispatching exactly the fix items; it must point at the review file, the plan, and the previous report. Re-review changes and affected dependents using prior evidence for unchanged scope; repeat the full matrix only when required by the agreed acceptance or blast radius. Separate source-semantic defects from verification-environment defects before ordering regeneration.
7. Fix rounds keep their own artifacts: the fix session writes <slug>.report-2.md (then -3, …), never overwriting an earlier report, and each round gets its own <slug>.review-2.md (…), so every round stays auditable.
8. Exit: a 符合 verdict closes the loop — record Status: done (rounds: N) at the top of the plan doc, list anything still not done, and update project memory/docs per project rules. If a round makes no progress on the same blocking item, stop and hand the open questions back to the user instead of looping again.

Courier note: the return leg is user-mediated — this session is not woken by the executor and resumes only when the user brings the report back. Say so when handing off, so nobody waits for an automatic ping.

Hard boundary

Once plan + launcher prompt are delivered, the planning job is DONE for now. Do not execute the plan here, do not spawn execution sub-agents, do not start a fix round yourself. The user launches the fresh session; this session's next role is reviewer (see Review loop).