Grill me.

The assistant never tells you when it guessed.

Handed a vague job, it does not stop and ask. It picks an answer silently, commits to it, and hands back something confident and wrong. Most bad AI output is not a bad model. It is an unanswered question that went unnoticed.

Grilling makes the guesses visible before they cost anything. The interview is the point; the skill file is just how you make it happen every time without retyping.

Save as grill-me/SKILL.md, or paste the body into a new chat.

---
name: grill-me
description: Interview me one question at a time until you know enough to do
  the job properly, then write the brief. Use this when I say "grill me",
  "pressure test this", "interview me first", "ask me before you start", or
  when I ask you to make sure you understand a job before doing it. Do not
  start an interview unless I have asked for one.
---

Before you do this job, find out what you do not know about it.

Interview me until you could hand the job to somebody else and they would get
it right.

1. One question at a time. Wait for my answer before the next one. Never stack
   three questions into one message.

2. With every question, give me your recommended answer and one sentence on
   why. I may not know the answer cold. Reacting to yours is easier than
   inventing mine.

3. Ask about anything you would otherwise have to guess: where this is going,
   who reads it, what they already know, what it should make them do, what I
   would never say. If you are about to assume something, that is a question.

4. Look it up before you ask. If the answer is in a file I gave you or in
   something I already said, go and find it. Ask me only about what is
   genuinely mine to decide.

5. Push back on vague answers. "More clients" is not an answer. "Six retainer
   clients by December" is. Ask for the number, the date, or the name.

6. Tell me what you would have gotten wrong. Each time I answer, say what you
   had been assuming before I told you.

7. Stop after about eight questions, or sooner if I say "that's enough."

Then write the brief:

- The job, in one sentence
- Who it is for, and what they already know
- What it has to include, and what it has to avoid
- What a good one looks like
- Anything you are still guessing at, marked clearly as a guess

Then stop. Do not do the job until I say go.

Rule 6 is the one to watch. It is what turns a normal back and forth into a lesson: the assistant says what it had wrong, in real time, in its own words.

Everything under the dashes is written for you. The description at the top is written for the assistant. It is not a summary and it is not documentation; it is how the assistant decides whether this skill is relevant right now. A skill with a vague description sits there and never fires, and the owner concludes skills do not work. Hence the list of phrases: you will not remember the exact wording next Thursday, so the skill accepts several. Say any of them and it starts.

The last sentence matters more than the list: do not start an interview unless I have asked for one.Without it, the sensible sounding version of this skill is “use whenever a request is vague,” which fires on nearly everything, because nearly every request is vague. Being interrogated when you wanted a quick answer gets old within a day, and then the skill gets deleted. Trigger on what the person said, never on what the job looks like. If it still fires when you did not want it, add disable-model-invocation: true under the description. Then it only ever runs when called by name.

Put it where skills go, or paste it in.

• Claude. Skills live under Settings, then Capabilities, in the Claude app and on claude.ai: add the file there. In Claude Code, save it as .claude/skills/grill-me/SKILL.md inside your project, or under ~/.claude/skills/ to have it everywhere.
• ChatGPT.There is no skills folder. Put the body (everything below the second row of dashes) in a project’s instructions, or in the instructions of a custom GPT, and it behaves the same way.
• Anywhere else. Open a new chat and paste the whole thing in. It works either way. The file location is a convenience, not the point.

If you cannot find where skills go in thirty seconds, take the paste route and move on.

Then say “grill me” above something real.

Pick a job you were about to hand over anyway: summarize a client call, write a description for a service, make a post about your new hours. Run it once plain. Then run it again with “grill me” in front, answer the questions, and compare the two results.

It did not ask any of that the first time. It just picked, and it did not tell you it picked. That is what this fixes. The brief that comes out at the end is not homework. It is the prompt you were going to write badly, written properly, and it can be pasted straight back in.

This one is mine and I use it constantly. Notice what is in it: I never tell it what my answer is. I tell it how to behave, and I tell it when to stop. That is most of what a skill is. If you want to write your own, the fastest way is to grill yourself on it with this one.