Grilling Skill: Interrogate Your Plan with a Design Tree

Skill Overview

Grilling is an interrogation-style interview skill that asks successive questions about your plan, decisions, or ideas until both parties have reached a complete understanding of the matter, and only then begins implementation.

Use Cases

1. Stress-testing a plan before writing a PRD or code — When the requirements are still at the “it’s roughly like this” stage, Grilling asks about each ambiguity one by one, preventing you from discovering only after implementation that you misunderstood the direction.
2. Clarifying your thinking before making major decisions — For decisions such as a startup direction, technology selection, or architectural changes, it proceeds from the root node of the design tree to the leaves, bringing to the surface assumptions you have taken for granted but never actually verified.
3. When an idea is still only a rough outline — If all you have is a vague one-sentence idea, it can expand the branches for you and ask about aspects you have not yet fully thought through yourself.

Core Features

1. Design tree and frontier — It models the topic as a design tree: each decision has decisions derived from it hanging below it. The “frontier” is the set of all questions that can be answered now because their prerequisites have been determined. Each round asks only about the current frontier; two mutually dependent questions will never appear in the same round.
2. Question format that supports answering an entire round — Each question has a number and a title, followed by its recommended answer on a separate line (➡️). You can simply reply, “1 Agree, 2 Choose the second option, 3 No, because…,” without quoting the questions again. Once you answer, the frontier advances outward, and the next round asks precisely about the content unlocked by the answers from this round.
3. It handles the facts; you make the decisions — It independently verifies facts that need to be checked in the environment, such as reading files, running commands, or delegating to sub-agents; it will not ask you to do so. Decisions, however, must wait for your approval. Making decisions on your behalf would defeat the purpose of the skill. When the frontier is empty, it also will not automatically start working; it waits for you to confirm that your understanding is aligned before proceeding.

One boundary to keep in mind: the frontier is the model’s own judgment, not a computed graph. It may place two questions that should have been asked in separate rounds into the same round, only to discover later that the answer to one would have changed the other. If this happens, point it out directly, and the affected branches will be reopened in the next round. There is also no limit on the number of questions—some plans need only three, while others require fifty. Saying “stop” in natural language is the intended way to end the process.

Frequently Asked Questions

What is the difference between Grilling, grill-me, and grill-with-docs?

Grilling is the underlying primitive in this family of skills. It only handles the interview itself and does not write any files. It is also the one most commonly invoked automatically by the model, so you rarely need to run it manually. grill-me is a one-line wrapper around it, intended for situations where you have no working directory but want the process to have a clear name. grill-with-docs is likewise a wrapper, but it also writes CONTEXT.md and an ADR while conducting the interview. Both require Grilling itself to be installed first; installing them alone will have no effect.

Can it be changed to ask only one question at a time?

Yes, and quite a few users work this way. Add a line such as “Ask only one question at a time during grilling” to your own CLAUDE.md or AGENTS.md. The default round-based questioning format is indeed controversial: people who read more slowly, work in a second language, or find a one-question-at-a-time rhythm helpful for maintaining focus generally report that this approach suits them better. This opt-out mode is officially supported.

What if it starts working on its own after asking the questions?

The confirmation step exists precisely for this situation: the skill does not end when the questions are finished; it ends only after you confirm that your understanding is aligned. Weaker models may still violate this by compressing “interview until mutual understanding” into a few questions followed by an outline. If this happens, the most reliable approach is to write “Do not begin implementation without permission” in your own AGENTS.md or CLAUDE.md. If it answers the questions itself, that is a runtime defect rather than the intended design—another reason the skill text separates facts from decisions.