Requirements to Spec

The most expensive bug is building the wrong thing correctly. Code is cheap now — AI generates it in minutes — which makes deciding what to build the bottleneck and the highest-leverage skill. A day of spec work routinely deletes a month of wrong code.

Gate 1: Find the job, not the feature

Every request arrives as a solution ("add an export button"). Dig to the job behind it before accepting it:

• Who exactly hits this, when, doing what? ("The accountant, every month-end, retyping invoices into Tally.")
• What happens today without the feature? The workaround is the spec's best teacher — and if there's no workaround, question whether the pain is real.
• What does failure cost? Annoyance, money, legal trouble? This sets the quality bar (a report can be eventually-consistent; a stock count cannot).
• Ask "what would make this a waste of time to build?" — the answer surfaces the hidden assumption that kills projects.

The requester's solution is a clue, not a contract. Sometimes the export button is right; often the job is "get data into Tally" and the right build is different.

Gate 2: Write the spec — one page, falsifiable

A spec is not a novel. One page, these sections, every line checkable:

1. Job statement — one sentence: who + situation + outcome.
2. Golden path — the numbered walk-through of the main flow, concrete ("Wholesaler opens Sales → taps New Invoice → …"). If you can't narrate it step by step, you don't understand it yet.
3. Acceptance criteria — bullet list of testable statements ("a payment exceeding balance is rejected with a visible error"). Each becomes a test later (testing-strategy). "Should be fast/easy/intuitive" is not a criterion until it has a number or an observable behavior.
4. Edge decisions made now — the 5–10 ugly questions answered in writing: empties, duplicates, permissions, concurrency, deletion, the boundary values. These are where "done" gets renegotiated later if you skip them.
5. Non-goals — explicitly listed. The non-goals section prevents more scope creep than any process. "V1 does NOT include: multi-currency, offline mode, bulk import."
6. Open questions — what you still don't know, with an owner. An honest unknown beats a silent guess.

Gate 3: Slice — walking skeleton first

• Cut scope by thinning every layer, never deleting a layer: version 1 is the full path (UI → API → DB → result) for ONE narrow case, ugly but real. A walking skeleton that works end-to-end teaches more than three polished-but-isolated components.
• Order slices by risk, not ease: build the part you're least sure about first (the tricky calculation, the third-party integration), because that's where the spec is most likely wrong and cheapest to fix early.
• Each slice independently shippable and demoable. "80% done" on ten things is 0 shipped; done on three things is a product.

Gate 4: The spec is a hypothesis — plan its test

• Name, before building, how you'll know the feature worked for the business: the metric or observable behavior ("support tickets about X drop", "clients stop exporting to Excel"). Building without this is decorating.
• Put the spec in front of the requester as scenarios, not abstractions: "So when a customer returns 3 of 10 items after partial payment, the system will ___ — correct?" Concrete scenarios extract corrections; abstract nods extract nothing.
• Expect the spec to be wrong somewhere. The plan must include a moment (after the skeleton, after first demo) where changing course is cheap and expected — not a failure.

Anti-patterns to refuse

• Coding to end the ambiguity. Ambiguity resolved by code is a guess with a large deletion cost. Resolve it with a question or a written assumption first.
• The everything-MVP. If the "minimum" takes three months, it's not minimum — re-run Gate 3.
• Spec by feature-list ("has invoices, has reports, has users") with no golden path — a list of nouns is not a design.
• Silent assumption absorption: an AI (or dev) filling spec gaps with plausible defaults and telling no one. Every assumption made while building gets written back into the spec and flagged to the requester (ai-build-quality Law 1).
• Polishing the spec forever. One page + the ugly questions answered + a skeleton slice is enough to start; the skeleton will teach you the rest.