grill-me: Turn a Vague Idea into an Actionable Plan Through Probing Questions

Skill Overview

grill-me is a stateless, AI-driven interview skill that repeatedly questions an unfinished idea, plan, or design in your head until you can genuinely make decisions about it.

Use Cases

1. When you have only a vague product idea and don’t know where to start. You don’t need to bring a finished plan to begin—the purpose of the session is to create that plan. The vaguer it is, the better suited it is, because ambiguity is exactly what it is designed to eliminate.
2. When you want to align your thinking before writing code, a proposal, or an article. It works for any idea worth taking seriously: a feature, a product direction, a business judgment, or an article. If you can already describe it precisely, you don’t need to be grilled.
3. When you worry that AI will simply go along with you and package underdeveloped ideas as certainty. Each round of grill-me provides its own recommended answers and forces you to state whether you agree or disagree, rather than having you nod forty times to produce a plan that was “written by AI and merely not opposed by you.”

Core Features

1. It asks questions in rounds according to a decision tree and won’t ask questions you cannot answer yet. It organizes the outstanding decisions into a decision tree. Each round asks only the questions whose prerequisites have already been established—that is, the current boundary—lists them with numbers, and includes recommended answers. After you respond, it recalculates the next round. As a result, later questions are always clearly grounded in your earlier answers.

2. It runs statelessly, writes no files, and leaves no workspace behind. It does not read the repository, produce CONTEXT.md or ADRs, or require the discussion topic to be related to code. When the session ends, the only thing left behind is a sharper version of the idea in your head. This lightweight nature is also what makes it useful anywhere.

3. It actively distinguishes between questions that can be resolved through discussion and those that cannot. Questions such as “one long document or three pages?” or “what should this interaction feel like?” cannot be answered through discussion alone—you need something to look at and interact with. When it encounters questions like these, it stops probing and asks you to first build a one-off prototype, then return with a one-line answer after seeing it. This prevents the session from expanding indefinitely.

Frequently Asked Questions

Is grill-me the same skill as grilling?

No, though they are closely related. grilling is the underlying questioning engine that defines the entire interview process; grill-me is its user-facing front door. It carries nothing itself and only hands you over to grilling for one round. Therefore, you must explicitly invoke grill-me with /grill-me; the AI will not pick it up on its own.

How many questions will a session ask? When is it finished?

Count rounds, not questions. Asking forty-six questions over four rounds is perfectly normal. The session ends when the boundaries have been cleared: every branch of the decision tree has been covered, and nothing has been silently assumed. If it asks you two hundred questions in one go, the scope is usually too broad—first have the AI break the work into smaller pieces, then grill each one separately. An overly long session can also drift into the “dumb zone,” where the context is too full and question quality drops noticeably.

What should I do after the session is complete?

Continue in the same session; do not start a new one. The value of the session lies in the context you have just established. If the result is indeed software, hand the same conversation over to to-spec to continue into the build process (this is optional, not the purpose of this skill). The division of labor between grill-me, grill-with-docs, and wayfinder is as follows: choose the former when you need to read a codebase and record conclusions in CONTEXT.md and ADRs; choose the latter when the work is too large to fit into a single session.

Usage Boundaries

The quality of grill-me’s output depends on the quality of your answers, not the number of questions asked. You need to take the lead: push back if the questions are less specific than you need; speak up if the scope starts drifting; and say “I don’t know” when you cannot answer—this usually means it is time to build a prototype rather than guess. At the same time, be careful not to stay in the interview too long without ever touching code. Also, don’t use it with plan mode enabled—plan mode urges the AI to produce a plan as quickly as possible, which is precisely the opposite of “stay in the questioning process.”