Don't Write Code Yet — Make the AI Interrogate You First: Socratic Requirements in Practice

The biggest waste isn't tokens — it's rebuilding the wrong thing

Watch enough vibe coding failures and a pattern emerges: the root cause is rarely "the AI can't write code." It's "you thought you were clear, and the AI thought it understood." Say "build me an expense tracker" — you're picturing a family shared ledger, the AI ships a single-user local app. Three rounds of rework later, the tokens are the least of it; what's really gone is your confidence.

My fix is counterintuitive: no code until the AI interrogates you. Not polite small talk — force it to ask at least 10 substantive questions, answer them, hear its summary back, and only then allow it to start. Stick with this habit and rework drops by at least half. Here's the full playbook.

The 10 questions: five dimensions of a requirement

You don't need all 10 every time, but none of these five dimensions may be skipped:

Users and context (2 questions): "Who uses this, and in what moment do they open it?" — turn "the user" from an abstraction into a concrete person in a concrete moment. "A commuter operating one-handed on the subway" and "an accountant at a desk" produce fundamentally different products.

Boundaries and non-goals (2): "What is explicitly out of scope for this version?" and "What input is absolutely forbidden?" — non-goals matter more than goals; they stop the AI from helpfully adding login, payments, and i18n on its own.

Data and state (2): "What does the core data look like? Give me three real examples." and "Where does data live, and what if it's lost?" — forcing real examples instead of abstractions is the litmus test for whether a requirement is real.

Success criteria (2): "Once built, how do I verify it's correct? Walk me through one concrete flow." — front-load acceptance so the AI walks the flow itself after coding instead of waiting for you to click around manually.

Constraints and order (2): "Any hard technical, budget, or time constraints?" and "If only one feature ships first, which one?" — the prioritization question cures "I want everything" and forces the true MVP cut.

One opener + the three-round convergence

A copy-paste opener:

"Before writing any code, ask me at least 10 questions to clarify the requirements. Make them specific — answerable with multiple choice or real examples, no correctly-useless questions. After I answer, restate the requirements in your own words; only start after I confirm."

Then converge in three rounds: round one, the AI asks and you answer ("I haven't decided — give me three options to pick from" is a valid answer); round two, the AI restates its understanding plus what it plans not to build, and you correct it; round three, the AI produces a one-page spec (feature list, data shapes, acceptance flow) for your sign-off. Rework falls off a cliff after that.

Three anti-patterns: when questioning fails anyway

Anti-pattern one: the AI asks correctly-useless questions. "Would you like the UI to look nice?" — reject it outright and tell it: only ask questions whose answers would change the implementation. That single instruction is the highest-leverage tuning of AI question quality.

Anti-pattern two: you answer for the AI. Too many "you decide"s hand over all decision power. The rule: every choice touching user value must be answered by you; only pure implementation details may be left to the AI.

Anti-pattern three: questions without restatement. Questions without a restatement round are the same as no questions — a human "mm-hmm" is as untrustworthy as an AI "got it." Forced restatement is the highest-ROI step in the whole flow; don't skip it.

When you may skip it: three exemptions

Not every task deserves 10 questions. Pure replication ("clone this screenshot"), one-off scripts ("batch-rename these 200 files"), exploratory spikes ("see if this library fits") — just start; over-questioning is its own waste. One test decides: if this task goes the wrong direction, is the rewrite expensive? If yes, interrogate first. If no, open fire.