TL;DR
Instead of guessing what context the AI needs, state your goal and command it to interview you — one question at a time — until it can do the job perfectly. One cheap round of clarification replaces the expensive loop of draft, miss, re-explain, and retry.
The Analogy
Walk into a tailor's shop and describe a suit from memory: "Dark blue. No — navy. Lapels, but not those wide ones. For a wedding. Maybe a funeral? Something breathable."
The tailor nods along, guesses, and sews. What comes back is a suit-shaped object that fits a guess. You'll wear it zero times.
Now walk into the same shop and say seven words: "I need a suit for a summer wedding." Then watch what the professional does. They don't sew. They measure. They ask about the venue — beach or ballroom? The month. Your role — guest, father of the bride, or photographer who'll be kneeling all afternoon? They feel the fabric drape, check how you stand, ask what shoes you'll wear. Only when the measuring is done does the scissors come out. The suit fits like a secret, because every question was a measurement you couldn't have remembered to volunteer.
Most people prompt AI like the first customer — front-loading guesses about what matters, hoping the pile of details is the right pile. Ask Me First turns you into the second customer: give the end goal, then let the expert run the measuring tape. The tailor always knew which questions mattered. You never did. That asymmetry is the entire technique.
The economics follow the same logic as the shop: a fitting costs the price of a conversation; a mis-sewn suit costs the suit. Clarifying questions are the cheapest tokens you will ever spend — and the follow-up corrections they prevent are the most expensive ones you'll ever make.
How It Works
The formula has three parts, all in the first message:
Goal + Interview command + Format
-
The goal — what "done" looks like, in one plain sentence. "Draft a LinkedIn post announcing my promotion to Engineering Manager." Not the details — the destination. The model needs to know what it's building before it can know what's missing.
-
The interview command — explicit permission and instruction to ask before doing: "Before you write anything, ask me up to 5 questions — one at a time — about whatever you need to know to make this excellent." This flips the model's default. Trained to deliver answers, it will happily produce a confident draft from thin air — unless you explicitly make interviewing the task.
-
The format — the shape of the final deliverable, exactly as in any good prompt: "Then deliver the post in under 150 words, warm but professional tone, no hashtags." The interview fills the container; the format defines it.
| Part | What it sounds like | Why it's there |
|---|---|---|
| Goal | "Draft a LinkedIn post announcing my promotion" | Defines "done" before details exist |
| Interview command | "Ask me up to 5 questions first, one at a time" | Makes information-gathering the task |
| Format | "Under 150 words, warm tone, no hashtags" | The container the answers will fill |
Notice what this technique really is: Ask Me First is how you gather the pieces for the perfect one-shot prompt. The interview is the context you couldn't write on your own — the tailor's questions are the specification you didn't know you needed. And the answers accumulate exactly where they should: on the context window, the desk the model re-reads before every reply. Five short Q&A exchanges cost a few hundred tokens of desk space and stay visible for the whole session; a single bloated "here's everything I can think of" brain-dump can cost more and still miss the one measurement that mattered.
Before & After (The Prompts)
Example 1 — the announcement
Write a LinkedIn post about my promotion to Engineering Manager. I'm excited, I've been at the company 4 years, I started as a junior, I like mentoring, my team is great, we use TypeScript and React, my manager's name is Sarah, and also there was a reorg last quarter. Also keep it humble but confident, not too long, engaging, professional.
The customer describing the suit from memory. Half of this pile is fabric the tailor would never choose — and the post that comes back reads exactly like this prompt: crowded, breathless, unsure what it's about.
I want a LinkedIn post announcing my promotion to Engineering Manager. Before writing it, ask me up to 5 questions — one at a time — about whatever you need to make this post land. Then deliver: under 150 words, warm professional tone, no hashtags.
The seven-word version. The model asks what you're most proud of — and suddenly the post has a spine instead of a list. It asks who you're addressing — peers, or the whole industry? — and the tone aims itself. Every answer is a measurement, not a guess.
Example 2 — the itinerary
Plan a 5-day trip to Japan for me. I like food, some culture, not too much walking, mid-budget, I've never been, and I don't like crowds.
You guessed six constraints. The planner needed to know: arriving where? Traveling with whom? First timer to Japan, or to travel? Any foods you won't eat? The wrong guesses become a beautiful itinerary for someone else's vacation.
Plan a 5-day trip to Japan. First, interview me — one question at a time, up to 5 questions — about everything you need to know. Only then build the day-by-day plan.
Now the gaps close before the planning starts. Day one doesn't include a 4 a.m. fish market if the fourth question reveals you're not a morning person.
Example 3 — the email
Write a polite but firm email to my landlord about the heating. It's been broken for 2 weeks, I've called twice, I mentioned it in person once, it's getting cold, I'm not sure about the legal side, I don't want to be aggressive, but it needs fixing.
The details are all there — but the strategy isn't. Firm how? Deadline stated or implied? Mention formal escalation or hold it in reserve? The draft guesses, and a guess about escalation is a dangerous guess.
I need to email my landlord about a heating outage entering its third week. Interview me first — one question at a time, up to 5 — about anything that changes how this email should be written. Then draft it: under 180 words, polite but firm.
The model asks the question you didn't know mattered — "Do you want to mention tenant rights now, or keep that as your next move?" — and the email gains leverage instead of volume.
Common Pitfalls
- Letting the AI ask twenty questions at once. Unconstrained, the interview arrives as a wall of interrogation — and a wall is a chore to answer, with half the questions irrelevant once the first answers land. Always cap the count ("up to 5") and demand one at a time: each answer changes what deserves asking next. That's what makes it a fitting, not a form.
- Answering the questions in advance anyway. If you append a paragraph of "context" to your Ask Me First prompt, you've described the suit from memory and asked for a measuring tape. Give the goal; let the questions do the digging.
- Deploying it on trivial tasks. "Ask me first" applied to translate this word to French is a tailor's fitting for a T-shirt off the rack. The technique pays for itself on deliverables — posts, emails, plans, documents — where a miss is expensive. Simple jobs want a simple one-shot prompt.
- Abandoning the interview halfway. Answering three questions and then typing "just write it" re-inserts guesswork at the exact moment it was about to be eliminated. If a question feels irrelevant, say why it's irrelevant — that's also an answer, and often the most informative one.
- Skipping the format. An interview with no target format ends with the model asking "how would you like this delivered?" — which is your job to have specified. Goal, interview, format: the formula is three parts, not two.
FAQ
What is reverse prompting?
Reverse prompting is flipping the direction of questions: instead of the user asking the AI, the AI asks the user — gathering the context it needs before producing anything. The "Ask Me First" technique is a structured form of it: you state the goal, command an interview (capped, one question at a time), and define the output format. The result is a first draft built from measured facts instead of guessed ones.
Should I give AI context upfront or let it ask?
Both — but in the right proportions. Give the goal upfront (one sentence, the destination), and let the interview surface the context. What you can articulate is rarely the full specification; the model's questions expose the dimensions you didn't know were missing. Rule of thumb: if you could write a complete one-shot prompt, do it. If you're unsure what matters, say the goal and open the interview.
Does asking questions first waste tokens?
No — it reallocates them. A five-question interview might cost a few hundred tokens; a missed draft plus two rounds of corrections can burn thousands and still end in a worse place. Clarification is cheap at the start of a conversation and expensive at the end. The only real waste is interviewing for a task that never needed it — which is why the technique is reserved for deliverables, not lookups.
Next Lesson
Ask Me First gets the facts right before the AI builds. But what about the answer you've already received — the draft, the plan, the argument that looks finished? There's a move that improves almost any output, no new information required: turn the machine against itself, and hire it as its own fiercest critic.
Continue to the next lesson: The Auditor Prompt
Practice drill before you go: pick the next real deliverable on your plate and open it with the three-part formula — goal, interview command, format. Answer only what it asks. Then compare the first draft to your usual first drafts, and count the correction rounds that never happened.