跳到正文
cccskills
免费在 GitHub 公开

grilling

The interview primitive for putting open decisions to an owner. Use on 'grill me', 'stress-test this', or 'what do you need from me before you build this', and whenever a plan, scope, spec, vision, or question ledger needs owner rulings. NOT for steps only a human can perform (use a guided session).

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(2)

  • SKILL.md5.6 KB
  • evals.md12.1 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

<!-- SYNCED COPY — do not edit here. Canonical: common-docs/skills/grilling/SKILL.md This file is distributed to every consuming repo by common-docs/meta/scripts/sync_skills.py. Edit the canonical, run the sync, and commit each repo. Edits made here are overwritten and lost. -->

grilling — interview by design tree

Mechanic adapted from Matt Pocock's grilling. Every sentence to the owner follows talk to Arman like a person; filing and delivery follow the ask-arman skill.

1. Build the tree, then prune it (before any question)

Draft privately: every open decision, and the decisions that hang off it. Then DELETE every node that is:

  • Settled — a node's DECISIONS.md or "settled — never re-ask" table, or table stakes (stream, persist, resume, never lose input = always yes).
  • A fact — dispatch a subagent (lane named). Never ask him anything you can look up.
  • Decidable from code, doctrine, or established best practice — decide it, with its companion machinery (decisions are complete). List it under "Decided" so he can override.
  • A number that is really a mechanism — bring the mechanism.

What survives is only: vision, product semantics, money, brand, legal, deleting real data, an account only he holds, his eyes on a real screen. The one exception: Arman invoking "grill me" on HIS OWN plan — then challenge his assumptions too, still never ask facts.

2. Rounds

  • Frontier = surviving nodes whose prerequisites are settled. A question whose answer depends on another question open this round waits for a later round.
  • Round size: per talk to Arman like a person; the rest wait on the Question Desk (ask-arman), ordered by how much of the tree each answer unblocks.
  • Never block on exploration. Send the fact-independent frontier now; only questions downstream of a running subagent wait. If he can step away while you explore, say so.
  • Number continuously across rounds (round 2 starts at Q5), so "5 yes" is never ambiguous.

3. Question shape — exactly this

Q<n> — <title>. 2–3 plain sentences of background: no doc references, codenames, item IDs, or emojis. Then the question, in one sentence.

  • Closed: options only where they differ in approach, never just in magnitude; the best practice in one sentence; Rec: one recommendation + the companion work that makes it true + why.
  • Open (vision): ask open, no rec; capture his words verbatim.
  • About a flow, surface, or behavior: attach the clickable URL + where to look, OR a mermaid diagram of what actually happens, OR a throwaway prototype (Artifact or demo route) labeled PROTOTYPE. A question he cannot answer from what you gave him is a defect in the question — and often a sign the PATH is broken, which is the real finding.

Every round ends with: Decided (override by number): D1 … · D2 … / Anything you skip ships with my recommendation. / What else should I know that I didn't ask? A question he skipped in an earlier round joins the Decided list as its own numbered item, carrying the recommendation that shipped and the words "default, not ruled" — never written as though he ruled it.

4. After every answer

  1. Record it where it belongs, then commit: vision → verbatim in VISION/STATE; ruling → the settled table, dated; creates work → the pending list or owning handoff; kills work → remove it.
  2. Recompute the frontier: prune branches the answer killed (say which queued questions died); unblock children; a partial answer gets a follow-up next round — never smoothed over.
  3. Skipped = the recommendation ships, recorded as "default, not ruled". "Defer" = a dated explicit deferral. "I need to see it first" = deferral + the URL he needs. "I don't know" / "we have options" = the question failed him: research, then bring it back next round with better facts, best practice, and a rec — never picked silently, never re-asked as-is (decisions are complete).

5. Done

The frontier is empty: every node is answered, decided-with-override-offered, or deferred with a date. Then act — no separate "confirm shared understanding" gate. State the final decisions in ≤10 lines in the same message in which you start the work.

Banned

One question per message · approval after each design section · "see the doc" · menus of numbers · asking whether it should stream/persist/resume · asking a fact · a fork with no recommendation · making him wait for a full audit before round 1 · inventing a question that was not on the pruned tree.

Never invent a question to have one to ask. A candidate that was not on the tree goes through the §1 prune first, like every other node — a decidable best practice spends a slot that belongs to vision. If the frontier comes up empty, say so and ship (§5); an empty frontier is the success condition, not a gap to fill.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Admin debug system wiring for the floating AdminIndicator's live debug data and Copy Full Context. Use when adding debug visibility to a route or feature, wiring useDebugContext or the debug panel, capturing console errors, or enabling the copy-context workflow for a page.

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

Compact two-icon Copy / Copy-for-AI controls (components/agent-copy). Use when adding copy buttons to a row, card, list, or record; merging duplicate Copy/JSON/AI controls; continuing the copy rollout; or writing a Copy-for-AI payload. NOT for markdown content actions (use rich-document-actions).

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

Disclosing a surface's existing fixed AI jobs in the shell's top Agents menu. Use when a page, panel, overlay, or window already runs a mandate behind a button, assist, automatic action, or mode; when the agent-disclosure guard names a file; or during a surface check.

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

The watch-fix-rerun method for making an existing platform agent or automated process efficient and correct: baseline its ledger, run one unit yourself, fix the class behind every wasted call, rerun, record. Use when asked to improve, optimize, watch, or 'make efficient' an agent, a sandbox session, a sync job, or any recurring automated process.

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

Redux state for agent execution and firing agent shortcuts. Use when editing features/agents/redux/, building agent UI, creating a conversation, touching assembleRequest or NDJSON stream state, wiring agent-state selectors, adding a per-conversation capability, or triggering a shortcut from a button, menu, or code (useShortcutTrigger, launchAgentExecution).

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

The provision for a call site: the exhaustive menu of values that place in the code can realistically produce. Use before creating or fixing a mandate or agent that reasons about more than its own input, when deciding whether an agent can answer at all with what it is sent, or when its output 'looks right' but nobody checked it could be. NOT for building the agent (use create-agent).

中文简介正在准备中。当前显示原文说明。

armanisadeghi/ai-matrx32026年10月20日 更新

查看 armanisadeghi 的全部技能

报告此技能的问题