菜单
Interrogate me about the task before any code — surface scope, constraints, edge cases, acceptance criteria, and unknowns through rounds of pointed questions until no material ambiguity remains. Use for "/grill-me", "grill me", "interrogate the requirements".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run one task through the plan→build→review pipeline — a Step-0.5 classify-task call routes the planner/implementer/reviewer models to the task's complexity tier, then the plan is written (planner subagent or inline), the implementer builds it in an isolated worktree, the my-review agent reviews the diff, and findings route by severity (low→issues, medium→triaged fix-list, high→collective replan, critical→own cycle). For single tasks not worth slicing into an issue graph. Use for "/pipeline <issue#|task>", "pipeline this".
Deep, security-weighted review of your changes or a PR — routes the reviewer model to the change's complexity (opus by default; fable for a complex diff, on ask), then spawns the my-review agent (xhigh reasoning) on that model. Read-only, report-only. Use for "/my-review", "/my-review <PR#>".
Classify one task or issue into a complexity tier — trivial, standard, or complex — and emit that tier plus a rationale; consumers resolve the tier's {model, effort} roster via the plugin's resolve-tier.sh. Grounds the call by fanning out 1–3 Explore subagents over the touched codebase areas, then asks you to confirm or override. Use for "/classify-task <issue#|brief>", "classify this task".
Run N rounds of the autonomous issue-solving loop — pick the ready set (blockers closed, skip hitl), fan out parallel implementers in isolated git worktrees, hand the completed branches to a merger that merges in dependency order and resolves conflicts under the done-check, close finished issues, then a reviewer files blocking follow-ups. Use for "/orchestrate", "run the loop", "build the ready issues".
Check whether the plan(s)/PRD(s)/issue-slice(s) under discussion still match the decisions made in this session — a sonnet subagent reads the current session log, later decisions win, and reports drift read-only. Use for "/verify-plan", "does my plan still match what we decided?".
Capture the just-approved plan after exiting plan mode and hand it to a fresh session — write the plan to a file plus the resume pointer the workflow plugin reads, then tell me to /clear and send `go` to execute it. Use for "/handoff-plan", "hand off this plan".
| name | grill-me |
| description | Interrogate me about the task before any code — surface scope, constraints, edge cases, acceptance criteria, and unknowns through rounds of pointed questions until no material ambiguity remains. Use for "/grill-me", "grill me", "interrogate the requirements". |
| argument-hint | [the task or feature to pin down] |
| model | inherit |
| effort | xhigh |
| allowed-tools | Read, Grep, Glob, AskUserQuestion |
Interrogate me about the task in $ARGUMENTS until we share a precise understanding of what to
build and how we'll know it's done. You ask; you do not write code or change files. Skim
enough of the repo (Glob/Grep/Read) to ask specific questions, then go round after round —
prefer AskUserQuestion when the answer is a choice between a few options. Ask as many questions
as you can think of, covering problem & why, scope boundaries, constraints (tech, perf, compat,
deadlines), edge cases & failure modes, acceptance criteria, and unknowns/risks. Each round
chases the gaps the last round's answers opened.
Don't stop early. Keep generating new questions until you genuinely can't think of another that would change the design — only then declare yourself done. Periodically restate what you've heard in a line or two so I can correct a misunderstanding before the end.
