Grill Me: Let AI ask you 50 questions before you write code

Failure mode: “The AI didn’t do what I wanted”

The first failure mode that Matt talked about in his speech is: you think that the requirements in your mind are very clear, and let the AI write them out - that is not the case at all.

"I would run it, and I would try not to look at the code, but I would look at the code, and I realized I would get worse code. I did it again, I got even worse code... I did it again, kept running the compiler, and I would just end up with garbage."

Many people are familiar with this feeling: if you say "Add a login for me", AI will not ask you "Do you want to remember the device?" "How many times have you failed to lock the account?" "How long does it take for the session to expire?" It directly lays out a plan that it thinks is reasonable. By the time you review it, 500 lines have been written - two hours of reworking.

Why is this happening: design concept deviates

Matt quotes Frederick P. Brooks’ design concept (design concept) in “The Design of Design”:

When multiple people collaborate to design something, there will be something being created between you - it is floating in your mind, an invisible "theory about this thing". It's not an asset, it's not an asset stuffed into a markdown file, it's an invisible consensus.

AI writes code as soon as it comes up, which means it doesn’t share the same design concept with you at all. What is wrong when writing code is not the syntax, but the premise.

To fix this problem, you must first do design concept alignment before starting. The tool Brooks gave is called design tree - split a decision into multiple branches, and then split each branch. You can't skip upstream decisions and make downstream decisions directly, otherwise everything will have to be redone once the upstream changes to the downstream.

Matt’s Skill full text

Matt's implementation of this theory in mattpocock/skills is productivity/grill-me/SKILL.md, and the entire file plus frontmatter is less than 15 lines:

---
name: grill-me
description: Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
---

Interview me relentlessly about every aspect of this plan until we reach
a shared understanding. Walk down each branch of the design tree,
resolving dependencies between decisions one-by-one. For each question,
provide your recommended answer.

Ask the questions one at a time.

If a question can be answered by exploring the codebase, explore the
codebase instead.

Interpret sentence by sentence:

• "interview me relentlessly" - The key word is relentlessly (not letting go). By default, LLM has a tendency to ask 1 or 2 questions and then feel "almost" and start taking action. This word forcibly suppresses this tendency.
• "walk down each branch of the design tree" —— Brooks' concept of design tree. Force Claude to treat your requirements as a tree, solving the upstream first and then the downstream. If you say "Login", it will first ask "Authentication method" (tree root), and then expand "How to manage session"/"How to save token" (sub-node) based on your answer.
• "resolving dependencies between decisions one-by-one" - Explicitly prohibit packaged questions. There are often dependencies between decisions (if you choose SSO, the downstream does not need password policy issues), first make sure that the upstream can eliminate a lot of downstream problems.
• "for each question, provide your recommended answer" - key bonus points. AI not only asks questions, but also recommends answers. You just nod/no and save 80% of typing time.
• "ask the questions one at a time" - Prevents the AI from giving you 10 questions at a time.
• "if a question can be answered by exploring the codebase, explore the codebase instead" - If it is a fact that already exists in the project (such as "What testing framework is used in the project"), let Claude see it himself, don't ask you.

7 lines, but each sentence corresponds to a specific LLM behavioral bias.

How to install and use

Installation:

npx skills@latest add mattpocock/skills

Check grill-me and setup-matt-pocock-skills (grill-me does not depend on the latter, but other skills depend on it, so it is recommended to install them together).

Call: Enter /grill-me in the Claude Code dialog.

Typical process:

1. You describe what you want to do, which can be very vague ("I want to add a comment function to my blog")
2. Enter /grill-me
3. Claude started asking questions one by one, and gave recommended answers to each question.
4. You answer each question one by one (nod/no/correct)
5. Generally, a consensus is reached after 20 to 50 questions, and Claude will give you a summary.
6. The summary can be directly fed to /to-prd to become PRD, or directly passed to /tdd to start writing.

Real case: How much does a video editor function cost?

Matt gave a few specific numbers in "5 Agent Skills I Use Every Day":

• New video editor feature - 16 questions to reach consensus
• Complex functions - 30~50 questions
• EXTREMELY COMPLEX - 100 questions, session up to 45 minutes

Sample question (restored from Matt's video/blog post):

• "Should video clips be reorderable, or only added/removed in sequence?"
• "When a clip is deleted, do we keep its source file, or delete the file too?"
• "Does the editor need undo/redo? How many steps deep?"
• "Should we render previews in the browser, or rely on a backend service?"

None of these issues were technical - they were all product decisions. But every decision determines the shape of hundreds of lines of code. If you skip these questions and let AI write them directly, it will come up with a set of answers on its own and you will come back to reject them one by one after writing them.

Differences between Plan Mode and Claude Code’s built-in Plan Mode

Claude Code comes with plan mode (press Shift+Tab to enter). On the surface, it looks similar to grill-me—discuss first before taking action. But Matt directly said in his speech that he prefers grill-me:

Specific differences:

The biggest practical difference is "urgent or not". Plan mode is in a hurry to start, grill-me is not in a hurry - it regards "thinking clearly" as the main task rather than the prologue.

Advanced usage

1. Non-programming scenarios

grill-me does not bind code and can also be used for pure product decision-making conversations. Matt uses it himself:

• Course syllabus design
• Article writing
• Internal communication documents

As long as you have a vague idea in your mind and want to be forced to think it through, you can use it.

2. Cooperate with /to-prd

After the grill-me session is over, just say /to-prd, and Claude will condense the entire conversation into a structured PRD (including user story, module splitting, and testing strategy) and submit it to your issue tracker. Key point: Do not clear context in the middle - to-prd is extracted directly from the conversation context and will not ask you again.

3. Cooperate with /grill-with-docs

If the project already has CONTEXT.md (domain language) and docs/adr/ (architectural decisions), use /grill-with-docs instead of /grill-me. It will synchronously update CONTEXT.md while being tortured - decisions are being made while documents are updated, and there is no longer the problem of "documents being forever out of date".

4. Customize question depth

If you are pressed for time, you can add a sentence directly after /grill-me: "Limit to 10 questions, focus only on architecture decisions." It will converge according to your limit. But Matt doesn’t recommend it – he believes that “asking more questions” is exactly the value of this skill, and cutting it off is almost like plan mode.

Notes

It will be annoying the first time you run. People who are used to "generating 500 lines in one sentence" will feel that it is a waste of time to be asked 30 times by the AI ​​for the first time. Matt’s advice is to hold on to the first 5 questions – the first 5 questions often reveal things you haven’t even thought about. Once you pass that threshold, you’ll be addicted.

Not suitable for extremely small tasks. Change a typo, add a console.log - don't use grill-me. It is suitable for "making a new thing" or "changing an old thing with side effects".

Sometimes the AI will ask for technical details. If you don't care and want to let it judge, just reply "your call" and "you decide", and it will accept it and continue.

Why is this Skill popular?

/grill-me is the most frequently screenshotted and forwarded skill among Matt’s skills. The reason is not complicated:

1. Extremely minimalist: 7 lines of markdown, just copy and paste
2. Immediate effect: You can feel the change in AI’s “problem density” during the first run.
3. Portable: Does not depend on Claude Code, Codex, Cursor, and Aider can all be used
4. Comes with anti-LLM default behavior: Each word is anti-LLM deviation, with high engineering aesthetics

Its success has also become the best argument for "skill does not necessarily have to be long".

Reference resources

Next article: Grill With Docs: Maintaining project language and ADR - an advanced version of grill-me, for projects with domain complexity.