CoT Question Cards (Chain of Thought Question Cards) is a public Agent Skill (SKILL.md) for plan alignment and design interview work. Skill id: question-cards (compat alias cot-question-cards). Companion for one-off A/B/C picks: choice-cards. Chinese product name: 闪电问.
Repository: https://github.com/StockOnLLC/cot-question-cards
Skill bodies
- Canonical loop:
skills/question-cards/SKILL.md - Compat alias:
skills/cot-question-cards/SKILL.md - Lightweight companion:
skills/choice-cards/SKILL.md
Teams searching for grill-me, grilling, grill-with-docs, AskUserQuestion, AskQuestion, request_user_input, choice-cards, Socratic interview, plan alignment, design interview, spec sharpening, product discovery, agentskills, and Agent Skills want structured cards that sharpen a plan without dumping half-baked schemes mid-loop.
CoT Question Cards keeps that card UX neighborhood while enforcing a V2 Master Loop:
- Fuzzy goal + samples / research — lock a rough effect and concrete analogs first
- Concrete macros — cover product surface, cadence, granularity, scope, bedrock, cost of being wrong
- Agent-private fission — expand 2–5 schemes, self-test, keep one champion draft
- Micro structure cards — refine load-bearing micro structure with the user
- Satisfied loop — after Optimized Solution ships; if unsatisfied → re-ask macros → micro again until fully satisfied
Outputs: Fuzzy Brief, Goal Contract, unique Optimized Solution, execution record, and a satisfied-loop regression trail.
| Phase | Name | Who answers | What ships |
|---|---|---|---|
| F | Fuzzy goal + samples/research | User (cards) | Fuzzy Brief (private) |
| 1 | Concrete macros | User (2–5 cards) | Goal Contract (private) |
| 2 | Agent fission | Agent only | Champion Draft |
| 3 | Micro structure | User (≤3/round) | Revised Optimized Solution + execute |
| S | Satisfied loop | User | Close, or re-macro → micro → satisfied again |
Plain sequence: fuzzy → samples or research → cover-all macros → private fission → micro cards → satisfied; if not satisfied, ask more macro, then micro, until satisfied.
- Satisfied cards never mix into the first fuzzy/macro decks; they appear only after there is delivered/executed output
- Mid-fission forbids multi-scheme user votes (fission = Agent-private; micro cards = after champion draft)
- Cards follow user language; Skill body is bilingual EN+ZH
- No question numbering with
①②③or Q1 - No leading
❓❗🚨(or similar) before the question text - Color marks 🟢🟡🔴🔵 define meaning only, unlimited count/ratio (multiple greens OK)
- Grok Bot must use native
type:widget; never fake clickable cards with markdown bullets - Option labels stay short; option descriptions are long enough to answer “选这个等于什么/会怎样”
| Old V1 | V2 |
|---|---|
| Start at dense macros | Start at fuzzy + samples/research |
| Micro never asked of the user | Micro asked after champion draft |
| Satisfied only once post-execute | Satisfied loops: unsatisfied → re-macro → micro → satisfied again |
| No explicit research fork | Research when the fuzzy goal mixes tool-kinds |
| Dimension | grill-me / grilling (typical) | AskUserQuestion / AskQuestion / request_user_input alone | choice-cards | CoT Question Cards |
|---|---|---|---|---|
| Primary job | Relentless interview | Collect answers via UI | One-off A/B/C | Plan/design alignment with V2 timing |
| Mid-fission votes | Often yes | Possible if author dumps schemes | N/A | Forbidden |
| Satisfied timing | Often early or ad hoc | Author-dependent | N/A | Satisfied loop after Optimized Solution |
| Product name | Grill family | Host primitive | Companion skill | CoT Question Cards/闪电问 |
CoT Question Cards uses the same host primitives (AskUserQuestion, AskQuestion, request_user_input, Grok widget, OpenCode question, Gemini ask_user, CoTree target-host primitives) while owning a distinct loop name, Goal Contract, unique Optimized Solution, and satisfied loop.
Always first. Lock a rough effect — what success should feel like — plus samples or research analogs. If the fuzzy goal mixes multiple tool-kinds, research public analogs, let the user pick which to lean on (multiple greens OK), then enter concrete macros. Skip research when the user already pasted samples. Forbidden in Phase F: satisfied cards, API/component/token-level detail, treating the fuzzy goal as an acceptance checklist.
Ask until aspects are covered: success criteria, audience/non-audience, cadence, granularity, bedrock constraints, cost of being wrong. Round size 2–5 load-bearing cards. Do not ask micro implementation yet. After answers, the Agent privately writes the Goal Contract (success/scope/bedrock/gates/sample refs) with no confirm card, then enters Phase 2.
From Goal Contract + samples: seed ≤3 levers → expand 2–5 schemes → self-test against gates → pick one champion draft. Forbidden: mid-fission “A or B?” preference polls.
Only after a champion draft exists. Ask load-bearing micro structure (information architecture, cadence triggers, granularity bands, key state machines) — ≤3 cards/round. If answers contradict the Goal Contract, jump back to Phase 1 macros. Revise the champion into the unique Optimized Solution and execute what can ship now.
Only after the Optimized Solution is delivered/executed:
- 🟢 Satisfied — adopt and close
- 🟡 Unsatisfied — say what is wrong → Agent returns to more macro → Phase 2 → Phase 3 → satisfied again
- 🔴 Macro/samples wrong → back to Phase F or Phase 1
- 🔵 Expand
Loop until fully satisfied. Two unsatisfied rounds that are actually fuzzy-goal wrong force Phase F.
| Host | Primitive |
|---|---|
| Grok Bot (desktop) | type:widget — prompt + options; 🟢→style:primary; 🔴→style:danger; allowCustom; multi-Q via steps |
| Grok Bot (iOS/Android) | Same widget API; title-only/blank option bodies = client render bug — use desktop or typed reply; do not switch to markdown fake buttons |
| Claude Code | AskUserQuestion |
| Cursor | AskQuestion (required when available in the session) |
| Codex | request_user_input |
| OpenCode | question |
| Gemini CLI | ask_user |
| CoTree/contree | Install into the target host and use that host’s primitive (CoTree itself does not render cards) |
| Other/missing primitive | Text card with color marks, labeled 非可点/non-clickable |
Color marks (meaning only, unlimited ratio):
| Mark | Meaning |
|---|---|
| 🟢 | Best/recommended |
| 🟡 | Acceptable/doable with cost |
| 🔴 | Risky/not recommended |
| 🔵 | Need expand — explain, then more cards |
Usually 3–5 substance options + 🔵 when useful + host allowCustom. No A| prefixes when the host already shows letters. Prefer swipe steps for multi-Q; keep Phases F/1/3/S in separate decks.
Folder name must equal frontmatter name: question-cards. New installs copy skills/question-cards. Compat path skills/cot-question-cards still resolves older installs. Companion: skills/choice-cards.
git clone https://github.com/StockOnLLC/cot-question-cards.git
cd cot-question-cards
# Claude Code / Coco
mkdir -p ~/.claude/skills && cp -R skills/question-cards ~/.claude/skills/question-cards
# Cursor (both roots recommended)
mkdir -p ~/.cursor/skills ~/.agents/skills
cp -R skills/question-cards ~/.cursor/skills/question-cards
cp -R skills/question-cards ~/.agents/skills/question-cards
# OpenAI Codex
mkdir -p ~/.codex/skills && cp -R skills/question-cards ~/.codex/skills/question-cards
# WorkBuddy / CodeBuddy
mkdir -p ~/.codebuddy/skills && cp -R skills/question-cards ~/.codebuddy/skills/question-cards
# OpenCode
mkdir -p ~/.opencode/skills && cp -R skills/question-cards ~/.opencode/skills/question-cards
# Gemini CLI
mkdir -p ~/.gemini/skills && cp -R skills/question-cards ~/.gemini/skills/question-cards
# Universal installer (when supported)
# npx skills add StockOnLLC/cot-question-cardsAlso works as an Agent Skill/SKILL.md on Windsurf, Cline, Continue, Trae, Zed, Copilot, Roo, Goose, Amp, Junie, AiderDesk, Cortex, Grok Bot, Kiro, Warp, Antigravity, and similar hosts that load SKILL.md folders. For CoTree/contree, copy the same folder into each host directory the contree spec points at (claude:~, codex:~, opencode:~, amp:~, cline:~). Web IDEs (Codespaces/Gitpod): commit .agents/skills/question-cards into the project, or paste SKILL.md into custom agent instructions.
Optional pointers in CLAUDE.md/AGENTS.md: load CoT Question Cards for plan/design alignment; do not dump the whole loop into every rules file.
Smoke test: Phase F/1 decks must not contain a satisfied option; after Optimized Solution execute, Phase S must appear; mid-fission must not present multi-scheme votes.
skills/
question-cards/SKILL.md # canonical CoT Question Cards V2
cot-question-cards/SKILL.md # thin compat alias → question-cards
choice-cards/SKILL.md # one-off A/B/C; prefer question-cards for trees
docs/brand/ # hero + loop + fission + layering + vs grill + IDE + 字标
LICENSE
README.md
## Fuzzy Brief/糊目标简报
- Rough effect: …
- Samples/analogs: …
## Goal Contract/目标契约
- Success/Scope/Bedrock/Gates: …
## Optimized Solution/最优方案(唯一项)
…
## Internal bake-off/比选摘要
- Variants: N; Passed: M; Winner: …
## Micro refinements/微观修订
…
## Execution/执行记录
→ What shipped: …
## Regression/满意循环
→ Satisfied close | Feedback → re-macro round k| Illegal | Why |
|---|---|
| Satisfied card inside Phase F/1 swipe | Breaks satisfied loop timing |
| Mid-fission “pick Scheme B or D” | Turns Agent into a waiter; no gates |
Numbering questions ①②③/Q1 |
Hurts card aesthetics/host layout |
Leading ❓ ❗ 🚨 on question text |
Noise; write the question sentence only |
| Markdown bullets as fake Grok cards | Must use type:widget |
| Forced one-green-one-yellow-one-red | Color marks are meaning-only; multiple greens OK |
| Mixing macro and micro in one deck | False precision before Goal Contract |
| Branding the product as Grill | Distinct name: CoT Question Cards/闪电问 |
CoT Question Cards sits in the Agent Skills/SKILL.md/agentskills/npx skills add neighborhood alongside searches for grill-me, grilling, AskUserQuestion, AskQuestion, request_user_input, choice-cards, Goal Contract, Optimized Solution, plan alignment, design interview, Claude Code skills, Cursor skills, Codex skills, OpenCode question, Gemini ask_user, CoTree contree, Grok widget, and Socratic product discovery. Related packaging names (anthropics/skills, Terminal Skills, Matt Pocock-style skill ecosystems) describe discovery channels — this repo owns the CoT Question Cards loop, not those brands.
Is CoT Question Cards just grill-me renamed?
No. Same card UX neighborhood; different V2 timing, Goal Contract, unique Optimized Solution, and satisfied loop.
Can I still use AskUserQuestion / AskQuestion / request_user_input?
Yes — as host primitives inside the V2 phases. CoT Question Cards is the timing discipline.
When is the Satisfied Card allowed?
Only in Phase S after Optimized Solution delivery/execution. Unsatisfied answers loop back to more macro, then micro, then satisfied again — not a single one-shot post-execute ask.
How does choice-cards relate?
One-off A/B/C. Decision trees and plan alignment → CoT Question Cards (question-cards).
Grok mobile shows title-only blank options?
Client render bug. Same type:widget API as desktop; ask the user to open on desktop or type the option text. Do not fake buttons in markdown.
Folder/id?
question-cards (frontmatter name must match). Compat: cot-question-cards.
| Term | Meaning |
|---|---|
| CoT Question Cards | English product name (this repo) |
| 闪电问 | Chinese product name |
| Question Cards/问题卡 | Skill body external name (never Grill) |
| Fuzzy Brief | Phase F private rough-effect + samples summary |
| Goal Contract | Phase 1 private success/scope/bedrock/gates artifact |
| Champion Draft | Phase 2 unique draft before micro |
| Optimized Solution | Unique champion after micro + execute |
| Satisfied loop | Phase S; unsatisfied → re-macro → micro → satisfied |
| choice-cards | Companion one-off pick skill |
| Agent Skills/SKILL.md | Packaging format |
Pull requests that clarify V2 timing, host install paths, Goal Contract examples, or platform primitive tables are welcome. Keep behavior changes in sync with skills/question-cards/SKILL.md. File timing bugs with host name, transcript snippet, and which phase showed a Satisfied Card or mid-fission vote too early.
MIT — see LICENSE. Copyright (c) 2026 StockOn.
- Repo: https://github.com/StockOnLLC/cot-question-cards
- Core:
skills/question-cards/SKILL.md - Compat:
skills/cot-question-cards/SKILL.md - Companion:
skills/choice-cards/SKILL.md
闪电问是 CoT Question Cards(Chain of Thought Question Cards)的中文产品名。Skill id:question-cards(兼容 cot-question-cards)。配套单次选择:choice-cards。
闪电问是一套公开 Agent Skill(SKILL.md),用来对齐计划或设计:先糊目标+样本/调研,再具象宏观,再由 Agent 内部裂变自测出冠军草案,再出微观结构卡,落地唯一最优方案后进入满意循环——不满意就回到更宏观重问,再微观,直到满意。
- 糊目标+样本/调研 — 大致效果 + 参照样本或 Agent 调研类似物
- 具象宏观 — 面面盖到(产品面、节奏、颗粒度、范围、基岩、做错代价)
- Agent 内裂变自测 — 扩 2–5 案、自测、只留冠军草案(禁止中途多方案投票)
- 微观结构卡 — 对冠军草案问承重微观结构(≤3 题/轮)
- 满意循环 — 有产出/执行之后才问;不满 → 宏观重问 → 微观 → 再满意,轮询到满意
- 满意卡禁止塞进第一轮糊目标/宏观滑组
- 裂变中途禁止半成品多方案让用户投票
- 题目前不加
①②③/Q1 - 题目前禁止
❓❗🚨 - 🟢🟡🔴🔵 只定含义,不限个数/配比(可三绿)
- Grok 必须
type:widget,禁止 markdown 子弹冒充可点卡 - 选项短 label + 长 description(选这个等于什么)
| 宿主 | 原语 |
|---|---|
| Grok Bot | type:widget(手机标题空白=客户端 bug,改桌面或打字) |
| Claude Code | AskUserQuestion |
| Cursor | AskQuestion |
| Codex | request_user_input |
| OpenCode | question |
| Gemini CLI | ask_user |
| CoTree | 装进目标宿主后用该宿主原语 |
git clone https://github.com/StockOnLLC/cot-question-cards.git
cd cot-question-cards
mkdir -p ~/.claude/skills && cp -R skills/question-cards ~/.claude/skills/question-cards
mkdir -p ~/.cursor/skills ~/.agents/skills
cp -R skills/question-cards ~/.cursor/skills/question-cards
cp -R skills/question-cards ~/.agents/skills/question-cards
mkdir -p ~/.codex/skills && cp -R skills/question-cards ~/.codex/skills/question-cards
# 亦可用:npx skills add StockOnLLC/cot-question-cards文件夹名必须等于 frontmatter name:question-cards。新安装优先 skills/question-cards;skills/cot-question-cards 仅为兼容别名。
卡片交互可以同类(AskUserQuestion/AskQuestion/request_user_input/Grok widget/OpenCode question/Gemini ask_user),但时序与产出不同:不叫 Grill;不中途投票;先糊目标与样本,再宏观,再裂变,再微观,落地后再进满意循环。英文产品名 CoT Question Cards,中文产品名 闪电问。
MIT — 见 LICENSE。







