Matt Pocock 新出的 /grill-with-docs：让 AI 真正理解你的项目语言

结论先行（BLUF）：TypeScript 圈最有影响力的技术布道者 Matt Pocock 把他用了一段时间的两个 Skill 合并成了一个——/grill-with-docs。它做一件事：在你开始写代码之前，逼 AI 用你项目的实际语言对你的设计方案进行严格审问，并实时把达成的共识写入文档。GitHub：mattpocock/skills（9K+ star）

---

背景：两个好 Skill，用起来有割裂感

Matt Pocock 之前发布了两个广受好评的 Skill：

• /grill-me：AI 扮演严格的审问者，针对你的方案逐一发问，暴露你没想清楚的地方
• /ubiquitous-language：建立项目专属词汇表（比如”materialization cascade”在你项目里是什么意思），让 AI 和你说同一种语言

问题是，两个 Skill 分开跑，流程不连贯：先建语言，再审问，文档还要自己另外更新。/grill-with-docs 把这三件事合成一个连贯动作。

---

/grill-with-docs 做了什么

公式：/grill-with-docs = /grill-me + /ubiquitous-language + 文档沉淀

具体流程：

① 读取 CONTEXT.md
这个文件记录项目的共享语言——业务里哪些词有固定含义、哪些概念不能混用。AI 先读懂这个上下文，再开始任何对话。

② 一次一个问题的严格审问
AI 不会一次甩出十个问题，而是一问一等，根据你的回答继续深挖。目标是暴露设计盲点、确认依赖关系，并在实际代码里验证你的假设是否成立，而不是空谈理论。

③ 术语冲突立即叫停
当你用的词和 CONTEXT.md 里的定义有出入，AI 会马上指出来，而不是默默猜测你的意思。如果遇到模糊表达，它会主动提议一个更精确的规范术语。

④ 实时更新 CONTEXT.md 和 ADR
每当对话中产生新的共识，AI 会当场把词汇定义更新进 CONTEXT.md，而不是等会话结束后再批量处理。如果某个决策足够重要（难以撤销、不显而易见、有真实取舍），就自动生成一条 ADR（Architecture Decision Record）。

---

为什么这件事重要

AI 写代码最大的失败模式不是代码质量差，而是写出来的东西和你实际想要的不一致。根本原因：AI 不知道你的项目在这个业务里说的”materialization”是什么意思。

/grill-with-docs 解决的是对齐问题，而不是代码生成问题。在动手写代码之前，先让 AI 真正理解项目的领域语言，之后所有的命名、结构和文档注释才会和现有代码库保持一致。

Oliver Ulvebne（@therealoliuliv）在用了之后说：一开始感觉被各种问题拖慢了，但用一阵子之后觉得它反而在省时间——因为后期的清理和返工少了很多。

---

快速安装

npx skills@latest add mattpocock/skills
# 选择 grill-with-docs，然后执行配置
/setup-matt-pocock-skills

在项目根目录创建 CONTEXT.md，写入你的业务术语表。第一次跑 /grill-with-docs 时 AI 会帮你把后续内容填充进去。

---

常见问题

Q：不用 CONTEXT.md 也能跑吗？
A：可以，但效果打折。没有 CONTEXT.md 时 AI 无法做术语校验，只是做通用的设计审问。第一次用时哪怕写几行领域词汇进去，效果差异就很明显。

Q：ADR 会自动创建吗，还是需要手动触发？
A：由 AI 判断。只有符合三个条件的决策才会被记录：难以撤销、对未来读者不显而易见、存在真实的方案取舍。不是每次对话都会生成 ADR。

Q：和直接问 Claude “你觉得这个方案怎么样”有什么区别？
A：结构化程度完全不同。临时提问得到的是即兴反馈，/grill-with-docs 得到的是基于你项目实际代码和既有语言约定的系统性审问，结论直接写入文档可供团队复用。

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

BLUF: TypeScript educator Matt Pocock merged two of his most popular skills into one: /grill-with-docs. It does one thing — before you write any code, forces the AI to rigorously question your design using your project’s actual domain language, while writing the agreed-upon decisions into your docs in real time. GitHub: mattpocock/skills (9K+ stars)

---

Background: Two Good Skills, One Awkward Workflow

Matt Pocock previously released two well-received skills:

• /grill-me: AI plays strict interrogator, asking pointed questions about your plan to surface gaps in your thinking
• /ubiquitous-language: Builds a project-specific vocabulary (e.g. what “materialization cascade” means in your codebase), so AI and developer share a common language

The problem: running them separately was disjointed. Build the language first, then audit, then update docs separately. /grill-with-docs collapses all three into one fluid session.

---

What /grill-with-docs Does

Formula: /grill-with-docs = /grill-me + /ubiquitous-language + live documentation

① Reads CONTEXT.md first
This file records the project’s shared language — fixed-meaning terms, concepts that can’t be conflated. The AI internalizes this context before any conversation begins.

② One question at a time
No ten-question dumps. One question, wait for the answer, then go deeper based on the response. The goal is to surface design blind spots and verify assumptions against actual code — not theoretical discussion.

③ Terminology conflicts flagged immediately
When your words diverge from CONTEXT.md definitions, the AI calls it out on the spot rather than silently guessing. Fuzzy language gets a proposed canonical term.

④ CONTEXT.md and ADRs updated inline
Every time the conversation produces consensus, the AI writes the new definition into CONTEXT.md immediately. Decisions that are hard to reverse, non-obvious, and represent genuine trade-offs automatically generate an ADR (Architecture Decision Record).

---

Why This Matters

The biggest failure mode in AI-assisted coding isn’t poor code quality — it’s misalignment. The AI doesn’t know what “materialization cascade” means in your specific domain. /grill-with-docs solves the alignment problem before the coding problem. Once the AI understands your domain language, all naming, structure, and inline comments will stay consistent with the existing codebase.

Oliver Ulvebne (@therealoliuliv) said after using it: it felt like it slowed him down at first with all the questions, but after using it a bit he honestly thinks it saves time overall — because the clean-up and polishing afterward shrinks dramatically.

---

Quick Start

npx skills@latest add mattpocock/skills
/setup-matt-pocock-skills

Create CONTEXT.md at your project root with a few domain terms. On the first /grill-with-docs run, the AI will help fill in the rest.

---

FAQ

Q: Does it work without CONTEXT.md?
A: Yes, but with reduced effectiveness. Without it, the AI skips terminology validation and just runs a generic design review. Even a few lines of domain vocabulary in CONTEXT.md makes a noticeable difference.

Q: Are ADRs created automatically or manually triggered?
A: AI-decided. Only decisions meeting three criteria get recorded: hard to reverse, non-obvious to future readers, and representing a genuine trade-off between alternatives. Not every session produces an ADR.

Q: How is this different from just asking Claude “what do you think of this plan?”
A: Entirely different structure. Ad-hoc feedback is improvised. /grill-with-docs produces systematic interrogation grounded in your actual codebase and established language conventions — with conclusions written to documentation for team reuse.

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.