菜单
Adversarial code review before a PR — Claude asks hard questions until you pass the test. Also supports elegant rewrites ("scrap this and implement the elegant solution") and spec-driven development. Use before opening any pull request. For mechanically finding bugs in a diff, prefer a platform-native code-review skill; this is the human-gate interrogation.
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
When the user wants to write, rewrite, or improve marketing copy for any page — homepage, landing, pricing, feature, about, or product pages; "headline help," "CTA copy," "value proposition," "tagline," "hero section," "make this more compelling," "conversion copy," or complete Figma-ready full-page copy (see Full-Page Mode below). For email copy, see email-sequence. For popups, see popup-cro. For editing existing copy, see copy-editing.
When the user wants to increase conversions on any marketing page — homepage, landing, pricing, feature, or blog; "CRO," "this page isn't converting," "low conversion rate," "bounce rate," "my landing page sucks" — or just shares a URL asking for feedback. For signup flows, see signup-flow-cro; post-signup activation, onboarding-cro; forms, form-cro; popups, popup-cro.
When the user wants to audit or diagnose SEO issues — "technical SEO," "why am I not ranking," "meta tags review," "my traffic dropped," "Google update hit me," "core web vitals," "crawl errors," "indexing issues," or even a vague "my SEO is bad." For pages at scale, see programmatic-seo. For structured data, see schema-markup. For AI search, see ai-seo.
Coder<->reviewer handoff loop (Claude<->Codex, roles swappable) with self-instrumented per-round cost tracking. Loops locally via HANDOFF.md until approved, then opens a PR with the full transcript.
Act as a Senior Accessibility Engineer to audit a page, component, or design against WCAG 2.2 AA. Produces a triaged finding list (critical/serious/moderate), focus-order map, contrast report, ARIA decision tree, and a screen-reader test plan. Use when shipping new UI, hardening an existing flow, or preparing for VPAT/accessibility conformance.
When the user wants to plan, design, or implement an A/B test or experiment — "split test," "variant copy," "multivariate test," "hypothesis," "statistical significance," "how long should I run this test," or comparing two versions. For tracking implementation, see analytics-tracking. For page-level conversion optimization, see page-cro.
| name | grill-me |
| description | Adversarial code review before a PR — Claude asks hard questions until you pass the test. Also supports elegant rewrites ("scrap this and implement the elegant solution") and spec-driven development. Use before opening any pull request. For mechanically finding bugs in a diff, prefer a platform-native code-review skill; this is the human-gate interrogation. |
| category | engineering |
| tier | core |
| slash_command | /grill |
| allowed-tools | Bash(git:*) Read Grep Glob |
Enter adversarial review mode. Challenge the current changes before a PR is created. Do not approve until satisfied.
/grill — adversarial review of current diff before /pr/grill rewrite — scrap and implement the elegant solution/grill spec — write detailed specs before handing off workPROJECT_ROOT=$(git rev-parse --show-toplevel)
cd "$PROJECT_ROOT"
git diff main
git diff --staged
Read the diff. Identify risks:
Ask the user 3–5 hard, specific questions about the changes. Do not accept vague answers — probe further. Only approve (and proceed with /pr) once satisfied.
After a messy fix, start fresh:
Ask clarifying questions until requirements are unambiguous. Produce a written spec (inputs, outputs, edge cases, acceptance criteria). Only begin implementation once spec is approved.
GATE: User has answered all questions satisfactorily. Only then proceed to /pr.