The Chris Voss method for AI: calibrated questions, labels and mirrors

Most of us brief an AI the way we send a WhatsApp message to a busy colleague: one message, then we wait. When the answer misses, we send a sharper command, then another. Each round costs time, and the AI never tells us what it was guessing.

Chris Voss, a former FBI hostage negotiator, learned that the most important information comes out only when you ask the right kind of question and show that you understood the answer. His tools work surprisingly well with AI. This guide shows two uses: briefing an AI through a short dialogue, and rehearsing your own difficult conversations.

✦

1

The idea in 30 seconds

Ask What or How first

"What would you need to know to do this well?"

Watch the 30-second explainer

The idea in 30 seconds0:00 / 0:31

1. Ask What or How first "What would you need to know to do this well?"
2. Label: "It seems like…" Name the drift you see, calmly and specifically.
3. Mirror vague words "A more robust approach?" makes the AI explain itself.
4. Start only after "That’s right" The AI sums up your goal; it starts when the summary is accurate.
5. Rehearse hard talks Let the AI play the other person, then coach you.
6. Ask, listen, check, then act.

Key takeaways

If you remember nothing else, remember these.

• Before the AI starts, ask open What or How questions so it writes its assumptions down where you can check them.
• Use labels to name drift and mirrors to make the AI explain a vague phrase.
• Hold a “That’s right” checkpoint: the AI starts only when its summary of your goal is accurate.
• Use AI as a rehearsal partner for appeals, authorship talks and supplier negotiations, without anyone’s personal data.

Who Chris Voss is, and why his method fits AI

Chris Voss led international kidnapping negotiations for the FBI. With writer Tahl Raz he wrote Never Split the Difference (2016), which turns crisis negotiation into everyday business techniques.

His central idea is tactical empathyUnderstanding the other side's position and feelings well enough to describe them accurately, out loud, before you try to persuade them.: understand the other side well enough to describe their view out loud before you try to persuade them.

The book is about people, not software. We borrow its tools because they solve a problem every AI user meets: the other party works from assumptions it never states. A negotiator draws those assumptions out with questions and summaries. You can do the same with an AI.

Six tools from the negotiation table

Here are the techniques this guide uses, paraphrased from the book. Tap a name for a short definition. The first four are the ones you will use with AI every week. The last two are mainly for people.

Voss also warns against “why” questions, because they sound like an accusation. With an AI there is another reason. Ask “Why did you do that?” and it will usually apologise and invent a plausible reason. “What assumption led to this?” gives you something you can act on.

Why this works with an AI

An AI does not get nervous and does not need to feel heard. These techniques work for a more practical reason: when your brief leaves a gap, the AI fills it with the most typical choice and carries on without telling you.

A calibrated question forces those gaps into the open. The AI writes its assumptions as text, and you can read and correct text in ten seconds. Compare the two versions below.

Before: vagueAfter: precise

Draft a 14-week teaching plan for Financial Accounting 2. The CLOs are attached.

Before you write anything: what would you need to know to do this well? List your assumptions, then wait.

What you typically get back

'I am assuming: no break week, one 2-hour lecture and one tutorial a week, equal CLO weighting. Questions: Which week is the mid-semester break? Is there a computer-lab session for the accounting software?' You fix three things before a single row is written.

A label does the same for drift: it names the pattern you see, so the AI adjusts that pattern instead of guessing what “better” means. A summary is a cheap test of understanding before the expensive part. None of this needs flattery. What matters is the order: ask, listen, check, then act.

Part 1: Brief the AI as a dialogue

Start with a normal AIM brief: Actor, Input, Mission. Then, instead of letting the AI run, add a short dialogue step.

Calibrated questions before it starts

Two questions do most of the work. “What would you need to know to do this well?” brings out missing information. “How would you approach this, and what could go wrong?” brings out the plan and its risks. Ask both, and tell the AI to wait.

You are a curriculum designer for Malaysian diploma programmes who works within OBE requirements.

Course: Financial Accounting 2, Diploma in Accountancy Year 2, 14 weeks, 4 contact hours a week. The three CLOs are pasted below. CLO2: Prepare a statement of cash flows using the indirect method.

Draft a week-by-week teaching plan as a table with columns Week | Topic | Activity | CLO.

Before you write anything: what would you need to know from me to do this well? How would you approach it, and what could go wrong? List your assumptions, ask me up to five questions, then wait for my answers.

Expect questions you had not thought of: whether students use accounting software in a lab, when the mid-semester break falls, how the CLOs are weighted. Each answer now is a correction you will not make later.

Labels to correct drift

When an output is off, the instinct is to complain or repeat the command louder. A label works better. Describe the pattern you see, then add a calibrated question.

Before: vagueAfter: precise

It seems like you're writing for completeness rather than for a busy dean who will read this on a phone.

What would you cut first if this had to fit on one page, and what must stay? Show me the plan before you rewrite.

What you typically get back

The AI names its trade-offs, proposes specific cuts you can approve, and protects the parts that matter. You decide before it rewrites.

Good labels are specific and neutral. “It seems like you’re treating every difference in the bank reconciliation as an error, when some are timing differences” is useful. “You’re being careless” is not, because it gives the AI nothing to adjust.

Mirrors for vague phrases

AI writing is full of phrases that sound precise and mean little: “a more robust approach”, “industry best practice”, “a holistic framework”. Copy the phrase back with a question mark: “A more robust approach?” The AI will usually explain what it meant, and you can judge whether it holds up.

One caution: some AI agentsAI agentAn AI that does a multi-step job for you (planning, opening files, using tools and checking its work) instead of only replying once. treat a short reply as an instruction and start changing things. If that happens, add: “Explain what you mean before you change anything.”

The “That’s right” checkpoint

Voss separates “You’re right”, which people often say to end a conversation, from “That’s right”, which they say when they feel accurately understood. Use that difference as a checkpointCheckpointA planned pause where the AI shows its work and waits for your OK.. The AI summarises your goal, you correct it, and it starts only when you would say “That’s right.”

Before you start, summarise in no more than five bullet points: my goal, who the output is for, the constraints you will respect, what "done" looks like, and anything you are still unsure about.

I will correct your summary. Do not begin the work until I reply "That's right."

This matters most with agents. An agent that misunderstands at step one repeats the mistake in every later step. A summary takes a minute to read. Repairing an afternoon of an agent’s work takes much longer.

For developers: technical detailsOptional: skip this if you only use the chat app

If you build agents, enforce the checkpoint in code as well as in the prompt. Run a planning step with read-only tools, show the plan to the user, and enable tools that write, send or delete only after an explicit approval. A prompt asks the model to wait; a permission gate makes sure it does.

Practise: rewrite questions and spot techniques

Try both rounds. In the first, turn closed and “why” questions into open How or What questions. In the second, name the technique each line uses.

Interactive

Calibrated question lab

1 · Convert it2 · Spot the technique

Pick the open question that starts with How or What.

Item 1 of 6

Score: 0

You, briefing an AI before a literature review

Closed question

Original: “Can you do a literature review on AI adoption in Malaysian SMEs?”

Are you able to do a good literature review?What would you need to know from me to do this review well, and how would you decide which papers to include?Why haven’t you asked me any questions yet?

Part 2: Rehearse real conversations with an AI partner

The second use has nothing to do with briefing. You ask the AI to play the other person in a difficult conversation, practise your lines, then ask it to step out of character and coach you.

It is a rehearsal, not a prediction. Real people will surprise you, but you will walk in having already practised a label, a mirror and two calibrated questions. Each card follows the same pattern: the AI’s role, the situation in general terms, the rules and how to coach you. Keep names and identifiers out.

A lecturer handling a grade appeal

Play a final-year student who is appealing a C+ on a group project report. You are polite but frustrated, and you believe one group member's poor contribution pulled the mark down.

My context: the rubric gave 20% to individual contribution, and peer-evaluation forms were collected. Our faculty has a formal appeal process. No names or student details are included here.

Run the conversation one turn at a time and stay in character until I type "pause". Then step out of character and coach me as a negotiation trainer.

When coaching: show where I used a mirror, a label or a calibrated question, where I argued instead of listened, and give me two better lines to try. Open by telling me why you came to see me.

Notice the power difference. You hold the grade, so your aim is understanding and a fair process, not “winning”. A good label (“It sounds like you feel the group mark hid your own work”) often lowers the temperature more than any explanation of the rubric.

A researcher negotiating authorship

Play a senior collaborator at another university. You expect to be first author because you secured the dataset, but you have contributed little to the analysis or writing. You are busy and slightly impatient.

My context: I ran the analysis and drafted most of the manuscript. We have no written authorship agreement yet. The submission deadline is in three weeks.

Stay in character until I type "pause". Push back realistically. Do not give in just because I ask nicely.

After I pause, coach me: which of my calibrated questions moved the conversation, and where would an accusation audit at the start have helped? Then suggest a fair agreement we could both sign.

An SME owner negotiating with a supplier

Play the sales manager of a regional ingredients distributor. You have just announced an 8% price increase on flour and butter, citing higher costs. You value this customer but have targets to meet.

My context: I run a small bakery in Ipoh and have bought from you for six years. My budget for these items is RM8,000 a month. I could commit to a 12-month order or accept a different delivery schedule.

Stay in character until I type "pause". Respond as a real supplier would: offer something only when I give you a reason to.

After I pause, list every calibrated question I asked, rate each one, and rewrite my weakest line as a "How" or "What" question.

Run each rehearsal two or three times. Then swap roles: you play the other person and let the AI demonstrate the techniques. Hearing a good label from the other side is a quick way to learn what one sounds like.

Use it honestly

Treat the AI’s coaching as a second opinion, not a verdict. It does not know your student, your collaborator or your supplier, and it can misread a situation with confidence.

Combine it with the other layers

The Voss loop does not replace a good brief. It starts from one. Write your AIM prompt first, then add a calibrated question and a “That’s right” checkpoint. If you name a method, such as the Chris Sanders method for an investigation, the dialogue checks that the AI understood it. The next guide, how to brief an AI agent, puts every layer together.

Check your understanding

Check yourself

The Voss method in five questions

1/5

Pick an answer, then tap Check answer. You can also use the arrow keys to move between answers.

Question 1 of 5: You are about to ask an AI for a 14-week teaching plan. Which line should you add first?

ACan you do this well?BWhat would you need to know from me to do this well? List your assumptions, then wait.CWhy do teaching plans usually go wrong?

Check answer

Nice work!

Lesson complete

You made it to the end. Take the quiz above if you have not yet, for bonus XP.

• 0 XP
• 0 day streak