菜单
Interview the user relentlessly about a plan or design until reaching shared understanding. Phase 2b gate of /development-lifecycle. Spawns 3 parallel reviewer hats (product/engineering/design) to stress-test the plan. Use when user wants to stress-test a plan, get grilled, or mentions "grill me".
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/redpanda-data/ui-harness --skill grill-me命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Snyk dependency sweeps across frontend, Go, and Bazel paths with npm false-positive triage and release-age gates.
Generate Codex hooks.json and AGENTS.md from Claude Code hooks. Maps supported hooks directly, including Edit|Write apply_patch aliases, and keeps Stop batch checks only as fallback. Use when setting up Codex compatibility or dual-agent support.
Question changed code as liability and remove unjustified surface area. Use before commit, push, PR, merge, or when a change feels overbuilt or low-value.
基于 SOC 职业分类
正在显示 SKILL.md
| name | grill-me |
| description | Interview the user relentlessly about a plan or design until reaching shared understanding. Phase 2b gate of /development-lifecycle. Spawns 3 parallel reviewer hats (product/engineering/design) to stress-test the plan. Use when user wants to stress-test a plan, get grilled, or mentions "grill me". |
Phase 2b gate. No implementation until plan survive grilling.
Walk every branch of decision tree. Resolve dependencies one-by-one. Each question, give recommended answer. One at time.
If question answerable by exploring codebase, explore first.
Once user present coherent plan, spawn three reviewers in parallel (single message, multiple Agent tool calls):
plan-product-hat: persona, pain, success metric, scope, reversibility, TTVplan-engineering-hat: architecture, error paths, perf, security, test strategy, rollbackplan-design-hat: flow, a11y, copy, visual consistency, states (empty/loading/error)Each emit {reviewer, status, findings[], must_answer[]} per findings-schema.md.
Consolidate all must_answer questions into single list, deduped. Surface to user. User answer each. Plan updated inline.
Any reviewer return status: BLOCKED -> plan no advance until blocking finding addressed or user override.
All hats return APPROVED or user override specific findings -> plan approved -> Phase 3 (Implement).
Skip three-hat fan-out only if:
Else, fan-out mandatory. [ETHOS: Grill Before Build]