Get work from AI you would put your name on.

Before you start

• One real piece of work due in the next week or two A memo, an analysis, a proposal, a plan, a hiring brief, some research. Something that matters. Not a quick email.
• What you would work from Notes, data exports, documents, past versions. Take out anything your company would not want pasted into an AI tool.
• One AI tool ChatGPT, Claude or Gemini. One that accepts file uploads helps. Use the same chat for Steps 2 to 4.
• 60 to 90 minutes Most of it is Step 4, doing the work. The two prompts take about half an hour together.

Tick these off as you gather them. They stay ticked on this device.

Why this works

Experts don't use better words. They use a better method.

When you ask an AI for a memo, it writes the average memo. Telling it to "act like a top consultant" changes the vocabulary and leaves the method alone. You get the same thinking, said with more confidence.

The best people in a field work differently. They start from raw material instead of someone else's summary. They use a real method: a table, a test, a comparison. And before they hand anything over, they check it against something outside their own head.

So this guide works on the brief, before any work starts. The first prompt rewrites your ask into the brief a top professional would work from. The second interviews you until every decision in that brief is yours, not a guess. Then the AI does the work, and something outside the chat tells you if it is right.

We'll follow one example the whole way through: a memo on whether to raise prices.

Step 1 5 min

Write the ask you would normally give

Write down the ask you would give an AI today, in your normal words. Don't improve it. This is your "before", and you will compare against it at the end.

Then gather what you would work from: the notes, the export, the document, last quarter's version. The better the material, the more the next step has to work with.

Step 2 10 min

Deepen the brief

Open a new chat. Copy the prompt, paste your ask and your material into the two slots, and send it.

I have a piece of work to do. Before you do any of it, rewrite my brief into the version the best professional in this field would work from. Deeper, not bigger.

MY TASK, as I would normally ask for it:
[paste your usual ask]

WHAT I HAVE TO WORK FROM:
[list or paste it: notes, data, documents, past versions]

Rules for this step:

* Do not do the task. No draft, no outline, no sample. This step only rewrites the brief.
* Do not role-play. "Write as a top consultant" changes the words, not the method. Tell me what that professional actually does differently.
* Deeper, not bigger. Do not add sections, deliverables or scope. Add rigour to the same piece of work.
* Be concrete to my task. Name actual sources, people and tools, not categories.

Work through these in order. Keep each answer short.

1. The task in one line, as I asked it. If it is a noun ("a report", "a deck"), add the verb behind it: what it has to achieve, for whom, under what constraint.
2. Which profession does this best, and what would a top person in it do here that I am not asking for?
3. What is the rawest material they would start from, instead of a summary? Say whether I seem to have it. If not, say exactly what is missing.
4. Which methods and tools would they use? Real ones I can use in this chat or a spreadsheet, not "best practices".
5. What, other than you or me, would show the result is right? A source document, the raw data, a test, a real person's answer. If nothing outside can check it, say so plainly. That is a risk.
6. One target where the obvious version is timid. Raise it sharply and say why it might be reachable. One only.

Then write the rewritten brief: the goal as a verb, the raw sources, the methods, the target and the outside check. Under 200 words.

Stop there. Do not start the work.

It comes back with six short answers and a rewritten brief. In the example, it looked like this:

Notice what changed. The memo became a decision with a deadline. "Look at the data" became a specific export. And there is now a way to tell whether the answer is right that doesn't depend on the AI.

Often the most useful line is the one that says what you are missing. Go and get it, or decide to work without it and accept the risk. Both are better than not knowing.

Step 3 15–20 min

Get grilled

A good brief still hides decisions only you can make. Stay in the same chat and send this:

Before you do the work, interview me until we agree on every decision in that brief. Nothing silently assumed.

Rules for this step:

* Ask in rounds. In each round, ask every question you can ask now without guessing at an answer I have not given yet. A question that depends on another open question waits for a later round.
* Number each question. Under each one, give your recommended answer, so I can reply "agree" or correct you.
* Facts are your job; decisions are mine. If the answer is in what I have pasted, find it yourself. Ask me only for decisions and for things only I know.
* After I answer, work out what my answers settled and what new questions they opened. Then ask the next round.
* Do not do the task during the interview. No drafts.
* Do not praise my answers.

When there is nothing left to decide, write the final brief with my answers in it, under 250 words. List anything I still need to fetch before you start. Then stop and wait for me to say "go".

It asks in rounds. Every question comes with its recommended answer, so most replies are one word: "agree". Correct it where it is wrong. Your corrections are the point.

Each round settles some decisions and opens others. It stops when nothing is left to decide, then writes the final brief and a list of anything you still need to fetch.

Step 4 30–45 min

Do the work

Fetch what the list asked for and paste or upload it. Then send this in the same chat:

Go. Do the work to the final brief.

Rules:

* Work only from the material I have given you in this chat. If you need something I have not given you, stop and ask. Do not fill the gap with a guess.
* Show the method, not just the answer: the table, the grouping, the sums, the sources you used.
* Mark anything you could not check as UNCHECKED.
* At the end, say plainly whether you hit the target in the brief. If you did not, say how close you got and what was missing.

This is where most of your time goes, and it should. Read the method, not just the conclusion. In the example, that means the table of usage bands and the break-even sums, not only the price at the bottom.

If it uses a number or a fact you never gave it, ask where it came from. An AI will fill a gap with something plausible unless it is told not to, and UNCHECKED is it being honest about that.

Step 5 10 min

Check it against something outside the AI

Now use the outside check from Step 2. Not the AI's opinion of its own work: the source document, last year's numbers, the five customer calls.

In the example, that meant checking what happened to cancellations after the last price change. If the memo said a 15% rise was safe for light users, but light users left in numbers last time, the memo is wrong, and you found out before the CEO did.

Then compare against your Step 1 ask. If you want to see the gap, paste that ask into a fresh chat with the same material. The difference is usually obvious in under a minute.

Keep the final brief. The next time the same kind of work comes round, you can start from it and skip most of Step 2.

Stuck?

Something went wrong and you can't tell which of the five it is. Copy this into the same chat, with the step and what came back pasted in:

I am using two prompts to get better work from an AI: one that rewrites my brief into the version a top professional would work from, and one that interviews me until every decision is mine. A step has gone wrong. The step is: [paste the step name]. Here is what I pasted and what came back: [paste]. Five failures are known for this method: the AI started doing the work before the brief was settled; the "expert" brief changed the words but not the method; deeper turned into bigger, with more sections instead of more rigour; it asked me for facts it could have found in what I gave it; and there is nothing outside the AI to check the result against. Tell me which one is happening, quote the line that shows it, and give me the one next action. Do not do the work and do not rewrite the brief.

Or go straight to the one that sounds like yours:

• It started doing the work during Step 2 or 3
• The "expert" brief is just fancier words
• Deeper turned into bigger
• It asks me things it could have found itself
• There is nothing outside the AI to check it against