The Brief-Back Prompt: Make Your AI Repeat the Assignment Before It Starts

Every good teacher I worked with had the same trick for the first five minutes of a project. Before anyone touched a pencil, one student had to explain the assignment back to the class. Not recite it. Explain it, in their own words, with the part they were least sure about said out loud.

Half the time, the explanation was wrong in a useful way. "So we write a poem about a river?" No — a poem from the point of view of a river. Two minutes of brief-back saved forty minutes of the wrong poem.

Language models make the river mistake constantly. They don't announce it. They just write the wrong poem, beautifully, and you discover the problem on paragraph four. Today's prompt borrows the classroom trick.

The prompt

Paste this at the end of any request that will take the model more than a minute to answer, or you more than a minute to read:

Before you start, brief me back.

1. In two sentences, restate what you think I'm asking for and who it's for.
2. List the assumptions you're making that I didn't state. Label each one "safe" or "risky".
3. Ask me up to three questions whose answers would most change the result.
4. Describe what the finished version will look like: length, format, tone,
   and one thing you will deliberately leave out.

Do not begin the task until I answer your questions or reply "go".

That last line matters more than it looks. Without it, many assistants will brief back and then barrel straight into the work anyway, which defeats the point.

Why it works

When you give a model a short request, it fills every gap with the most statistically ordinary choice. Ordinary audience, ordinary length, ordinary tone. Each of those silent choices is a fork in the road, and you only find out which path it took when you read the result.

Brief-back turns the forks into a list you can scan in fifteen seconds. The "risky" label is the secret ingredient: it forces the model to rank its own guesses, and the risky ones are almost always where your real intent lives.

A worked example

Say you type: "Write a one-page plan for our community garden's spring fundraiser." Here's the kind of brief-back you'll get:

• Restatement: A one-page plan for a fundraiser run by a volunteer community garden, aimed at the organizing committee.
• Assumptions: One page means roughly 400 words (safe). The budget for the event is under $500 (risky). The audience for the fundraiser is existing members, not the wider neighborhood (risky). It happens on a weekend (safe).
• Questions: What's the fundraising target? Has the garden run an event before? Is there anything the committee has already ruled out?
• Finished version: A table of tasks with owners and dates, a short budget, and a timeline. I'll leave out sponsorship outreach unless you want it.

Look at those two risky assumptions. If your garden is trying to win over the neighborhood, the whole plan changes — different venue, different pitch, different budget. You'd never have spotted that from a confident first draft. You spot it instantly from a list.

When to skip it

Brief-back is a speed bump, and speed bumps are annoying in the wrong place. Skip it when:

• The task is tiny ("rewrite this sentence shorter").
• You've already pasted a detailed spec and you're confident in it.
• You're exploring and actually want the model's default guess, so you can react to it.

Three variations worth stealing

The risky-only version. For people who find the full brief-back too chatty: "Before you start, list only the risky assumptions you're making. One line each."

The team version. Run the brief-back, then paste it into the ticket or doc before anyone starts. It doubles as a shared definition of done, and nobody has to write one from scratch.

The skeptic version. "Brief me back as a skeptical editor would: what's unclear, what's contradictory, and what am I probably not telling you?" This one stings a little. That's the point.

Today's homework

Pick the next request you'd normally fire off in one line. Add the brief-back block. Count the risky assumptions. If there are zero, your prompt was already good — congratulations. If there are three, you just saved yourself a rewrite.

Then reply "go," and watch how much less editing the answer needs.