烤我
- 作者仓库星标 0
- 作者仓库 claude-code-expert
- 领域
- 工程开发
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @reedmayhew18 · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- Docker
- 运行环境要求
- Docker
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
第一阶段:大局观
用户想要构建什么?它是给谁用的?成功是什么样的?
第 2 阶段:设计树
对于每个主要组件:有哪些选项?有哪些权衡?哪个选项符合约束条件?这个决定对其他决定意味着什么?
第 3 阶段:边缘情况和故障模式
当出现问题时会发生什么?性能限制是什么?安全考虑有哪些?需要什么数据验证?
第四阶段:合成
回答完所有问题后: 总结所做的所有决定 突出任何紧张或权衡 提出具体的实施计划 询问是否遗漏了什么
第一阶段:快速上下文
你在建造什么? (一句话)它需要做的最重要的事情是什么?有什么硬性约束吗? (技术堆栈、时间线、平台)
第 2 阶段:仅限关键决策
询问会阻碍进展的 3-5 个最大决策 跳过理论上的极端情况 - 专注于“我需要知道什么才能开始构建?”
第三阶段:快速总结
用要点总结决策 记下推迟到以后的任何内容 推荐下一步(通常为`/wizard`或`/plan-and-spec`) 检查点:在推荐后续步骤之前向用户呈现摘要。问:“这看起来合适吗?准备好继续前进,或者需要调整什么吗?”
# Grill Me - Design Tree Exploration
## Goal
Reach a shared, unambiguous understanding of what the user wants to build before any implementation begins. Success = a concrete implementation plan the user has approved, with all blocking decisions resolved.
## Dependencies
- Tools: Read, Grep, Glob (read-only — this skill gathers information, never modifies files)
- No external services or MCP servers required
## Context
This skill reads the existing project (CLAUDE.md, README, config files) to pre-fill known answers before asking any questions. The less the user has to repeat themselves, the better. Reference `reference/02-best-practices.md` if uncertain about Claude Code patterns.
---
Interview the user about their plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one.
## Modes
### Full Mode (default): `/grill-me [topic]`
Relentless, thorough interview. Explores every branch of the design tree until zero ambiguity remains. Use for complex features, architecture decisions, or anything where getting it wrong is expensive.
### Light Mode: `/grill-me lightly [topic]`
Quick, focused interview — 5-8 questions max. Gets the essential decisions made without deep-diving every branch. Use for quick setups, small features, or when the user just needs to fill in a few gaps. Skip edge cases and failure modes unless they're critical. Aim for "good enough to start" not "perfectly specified."
## Pre-Fill: Check the Codebase First
**Before asking ANY questions**, check if you're in an existing project:
1. Look for: CLAUDE.md, README.md, package.json, pyproject.toml, Cargo.toml, go.mod, Makefile, Dockerfile, src/, tests/
2. If an existing project is found:
- Read CLAUDE.md, README, and config files to understand the project
- Identify: tech stack, architecture, conventions, existing patterns
- Note what you already know and what's still unclear
- **Only ask about the gaps** — don't ask questions the codebase already answers
- Start by saying: "I've looked at your project. Here's what I understand: [summary]. Let me ask about what I'm less sure about..."
3. If no project exists (empty directory or new project):
- Start from scratch with the full interview
## Rules
1. If a question can be answered by exploring the codebase, explore the codebase instead of asking
2. Ask one focused question at a time, not batches
3. When a decision opens new branches (e.g., "advanced search" → filters, sorting, pagination), explore each branch (full mode) or note it for later (light mode)
4. Don't accept vague answers — ask follow-ups until the answer is specific and implementable
5. Track decisions made so far to avoid re-asking
## Process (Full Mode)
### Phase 1: Big Picture
- What is the user trying to build?
- Who is it for?
- What does success look like?
### Phase 2: Design Tree
For each major component:
- What are the options?
- What are the tradeoffs?
- Which option fits the constraints?
- What does this decision imply for other decisions?
### Phase 3: Edge Cases & Failure Modes
- What happens when things go wrong?
- What are the performance constraints?
- What are the security considerations?
- What data validation is needed?
### Phase 4: Synthesis
After all questions are answered:
1. Summarize all decisions made
2. Highlight any tensions or tradeoffs
3. Propose a concrete implementation plan
4. Ask if anything was missed
**CHECKPOINT:** Present the full synthesis to the user. Do NOT recommend next steps until they confirm.
Ask: "Does this capture everything accurately? Anything missing or wrong before we move to implementation?"
## Process (Light Mode)
### Phase 1: Quick Context
- What are you building? (one sentence)
- What's the most important thing it needs to do?
- Any hard constraints? (tech stack, timeline, platform)
### Phase 2: Key Decisions Only
- Ask about the 3-5 biggest decisions that would block progress
- Skip theoretical edge cases — focus on "what do I need to know to start building?"
### Phase 3: Quick Summary
1. Summarize decisions in bullet points
2. Note anything deferred for later
3. Recommend next step (usually `/wizard` or `/plan-and-spec`)
**CHECKPOINT:** Present the summary to the user before recommending next steps.
Ask: "Does this look right? Ready to move forward, or anything to adjust?"
## Output
- **Format:** Inline chat — a structured summary presented in the conversation, not saved to a file
- **Full mode deliverable:** Decisions list + tensions/tradeoffs + concrete implementation plan
- **Light mode deliverable:** Bullet-point decisions + deferred items + recommended next step
- **Save location:** None by default. If the user wants it persisted, write to `PLAN.md` or `PROGRESS.md` in the project root and note it survives context compaction.
## Key Principle
The goal is NOT to ask a fixed list of questions. The goal is to dynamically explore the design tree until the level of clarity matches the mode: **Full mode = zero ambiguity. Light mode = enough to start.**
章节与文件索引
按说明文件中的顺序排列。
目标
依赖关系
背景
完整模式(默认):/grill-me [主题]
灯光模式:/grill-me lightly [主题]
文件/命令 -> reference/02-best-practices.md · /grill-me [topic] · /grill-me lightly [topic] · /wizard · /plan-and-spec · PLAN.md · PROGRESS.md
内容校验值 -> 34318576ca9b
当前原文的自动摘译,尚未人工审校。
目标
在任何实施开始之前,对用户想要构建的内容达成共同的、明确的理解。成功 = 用户已批准的具体实施计划,并解决了所有阻塞决策。
依赖关系
工具:Read、Grep、Glob(只读 - 此技能收集信息,从不修改文件)不需要外部服务或 MCP 服务器
背景
此技能会读取现有项目(CLAUDE.md、README、配置文件),在提出任何问题之前预先填写已知答案。用户必须重复的内容越少越好。如果不确定 Claude 代码模式,请参考
reference/02-best-practices.md。采访用户了解他们的计划,直到我们达成共识。沿着设计树的每个分支走下去,一一解决决策之间的依赖关系。完整模式(默认):/grill-me [主题]
坚持不懈、彻底的采访。探索设计树的每个分支,直到模糊性为零。用于复杂的功能、架构决策或任何犯错代价高昂的事情。
灯光模式:/grill-me lightly [主题]
快速、集中的面试——最多 5-8 个问题。无需深入研究每个分支即可做出重要决策。用于快速设置、小功能或当用户只需要填补一些空白时。跳过边缘情况和故障模式,除非它们很关键。目标是“足够好,可以开始”,而不是“完全指定”。