Stop Your AI From Guessing: Get Grilled Before You Build

You do not have a communication problem with your AI. You have no interview.

It is 6:40 p.m. In Manila. Divina runs a five-person design studio out of a co-working space near Makati. She has just told Claude to "set up an automation that onboards new clients so I stop doing it by hand."

Ten minutes later, Claude hands her a slick intake form. Name, email, budget, project type. Clean. Professional. Completely wrong.

It does not ask what happens when a client picks "branding" and "website" in the same project. It does not ask what counts as a rush job. It does not ask who gets notified when a lead has no budget listed at all. Divina finds out the hard way, three clients in, when the automation quietly mis-routes a $4,000 project to her junior designer's inbox.

Nothing was broken. The agent did exactly what it was told. The problem is that "set up an automation that onboards new clients" was never enough information to build the right thing. It was enough information to build a thing.

Why does your AI agent keep building the wrong thing?

Because you are the bottleneck, not the model.

Every AI agent you have ever used, from Claude to ChatGPT to Codex, will build with whatever you give it. Give it two sentences, and it fills fifty gaps with guesses. Some guesses are right. Most, in a real business with real edge cases, are not.

You would not let a new hire build your client-onboarding process from a two-line Slack message. You would sit them down, ask what happens in the weird cases, and only then let them touch anything. Somehow, with AI, that step keeps getting skipped.

The tweet that made "grilling" the word of the week

On July 26, developer educator Matt Pocock, the person behind Total TypeScript and a former Vercel engineer with 315,000 followers, posted four words: "Get grilled before you build."

Within hours it had over 300 likes and 45 replies from builders arguing about when the habit is worth the friction. He followed it with the line that explains the whole idea in one breath: "Talk shop before you slop."

Pocock did not invent the underlying idea. He builds and publishes a small library of reusable AI agent instructions, and two of the most-used ones are called grill-me and grill-with-docs. The names are blunt on purpose. Before the agent is allowed to build anything, it has to grill you.

What does "grilling" an AI agent actually mean?

It means you tell the agent, in plain language, to interview you before it touches a single workflow. Not a wall of twenty questions dumped on you at once. A real interview, one question at a time.

The published version of the technique runs on five rules, and every one of them is doing real work:

1. Ask one question at a time. Not a checklist. A conversation.
2. Always offer a recommended answer. You are not filling out a form, you are correcting a first guess, which is faster for almost everyone.
3. Separate fact-finding from decision-making. The agent digs up what it can on its own. You only get asked about things it genuinely cannot know.
4. Require a shared-understanding check before anything gets built. No automation goes live until you have both said, in effect, "yes, that's it."
5. Stay stateless. Each grilling session starts clean, so old assumptions do not quietly leak into a new project.

Strip away the developer jargon and it is a habit any business owner already understands: ask before you assume, confirm before you commit, and never build the whole thing on a guess.

A worked example: grilling before you onboard a new client

Here is what it looks like when Divina tries again, this time telling Claude to grill her first instead of build first.

Claude: "When a lead selects more than one service, like branding and a website, should that count as one project or two separate ones? My guess: one project, tagged with both services."

Divina: "One project. Good guess."

Claude: "If a lead leaves the budget field blank, should the form still submit, or should it require a number first? My guess: let it submit, but flag it for you to review personally instead of auto-assigning it."

Divina: "Yes, exactly that. I almost lost a good client because a junior designer auto-replied to one of those with a lowball estimate."

Claude: "Got it. Last one: what makes a project 'rush'? Under two weeks, or does the client have to say the word 'urgent' themselves?"

Divina: "Under two weeks, automatically. Don't wait for them to say it."

Three questions. Maybe four minutes. And the automation that finally got built handles the exact three situations that would have quietly cost Divina money, a client, or both. None of those three edge cases were things she would have thought to write into a brief. They only surfaced because something was finally asking.

Isn't this just a longer prompt?

No, and the difference matters. A longer prompt is you guessing what the agent needs to know, up front, alone. Grilling flips who does the guessing. The agent proposes an answer based on what similar businesses usually do, and you only have to correct it when it's wrong.

That is a much smaller cognitive load than writing the perfect brief from scratch. You are not staring at a blank page trying to anticipate every edge case in your own business. You are reacting to a stream of specific, answerable questions, one at a time, each with a sensible default already attached.

Why this matters more now than it did a year ago

A year ago, an AI agent that misunderstood your brief cost you a wasted afternoon. Today, agents like Claude Code and Codex can run unattended for hours, wiring up entire workflows, sending draft messages, touching real customer data, before you ever look at the result again.

The faster an agent can build, the faster it can build the wrong thing at scale. A misunderstood brief used to mean a bad first draft. Now it can mean a client-facing form that has been silently mis-routing leads for three days, or a support agent that has been giving five customers a day the wrong return policy, before anyone notices.

Speed without an interview is not efficiency. It's just a bigger mess, delivered faster.

The one grilling question that saves a rebuild, by workflow

Grilling is not just for client onboarding. It is worth two minutes before you automate anything that touches money, customers, or stock. Here is the single question that tends to matter most in five common small-business workflows.

How do I add a "grill me first" step to my own AI workflow this week?

You do not need to install anything or read a line of code. This works in Claude, ChatGPT, or Codex, in a plain chat window.

1. Open a fresh chat with whatever AI tool you already use for building or automating things.
2. Describe the workflow you want in one or two sentences, exactly as you normally would.
3. Add the grilling instruction before you hit send: "Before you build anything, interview me about this. Ask one question at a time, always suggest a recommended answer, and don't start building until we both agree we understand the plan the same way."
4. Answer honestly, and push back when a recommended answer is wrong for your business. That correction is the whole point.
5. Only once you get a summary you agree with, say so explicitly, something like "yes, that's it, go ahead and build it."

That is the entire technique. No plugin, no repository, no engineering background required. Just a habit of insisting on the interview before the build, every single time.

What you have after this becomes a habit

A week in, you catch the first misunderstanding before it ships instead of after a client complains.

A month in, every new automation you build starts with a two-minute conversation instead of a guess, and you stop tearing down half-built workflows because the agent assumed something you never said out loud.

Six months in, your team builds this way by default. Nobody hands an agent a two-line brief and walks away anymore, because everyone has felt what it costs when they used to.

Same tools. Same AI. A completely different hit rate.