The one-pager and the PRD
Write a one-pager that gets a yes or no from one person in five minutes, then a PRD in which every section carries the number the team builds to.
On this page
A one-pager exists to get one decision from one person; a PRD exists so that a team of six builds the same thing. They fail in opposite directions. A one-pager that tries to specify never reaches its ask, and a PRD that tries to persuade gets six different builds, because persuasion is written in adjectives and every reader fills an adjective with their own number. In both documents, the section without a number is the one the meeting argues about.
Two documents, two readers
The one-pager's reader is the person who decides — a head of product, a general manager — and they give it about five minutes. Its output is one of three words: yes, no or not yet. The PRD's reader is the team: engineers, a designer, the FDE, the support lead whose agents will use the feature. Its output is a build, and it lives as long as the build does.
The order is fixed. A PRD written before the decision is a sunk cost that argues for itself in the room, so the one-pager comes first; the PRD follows a yes, and its first section is the one-pager's numbers copied verbatim, because those are what was approved. A "not yet" sends the work back to discovery with a named test, which is what the one-day prototype is for.
The FDE's scoping doc is the same family for one customer: a PRD in content, a one-pager in discipline, two pages with a signature at the bottom.
The one-pager that gets no decision
The first draft for the drafting feature read like this:
Background. Support teams are under pressure, and AI is changing how they work. Vision. An assistant at every agent's side. Proposal. Build AI drafting into the console. Next steps. Discuss at planning.
Four sections and no number. The decider cannot say yes, because nothing in it can be held to account later; cannot say no, because nothing in it is wrong; and so says "let's discuss", which is the outcome the document existed to avoid: "discuss at planning" is a meeting, not a decision.
The one-pager, section by section
Six sections, each carrying the number that makes it decidable. Every figure comes from an earlier module: a one-pager summarises work already done.
The problem, as a count. The pilot tenant's 14 agents carry 130 tickets overnight every working day, and a reply takes 3 minutes — the queue page and the stopwatch from jobs to be done, not a survey.
The evidence, with its sample size. Forty interviews this quarter, and the prototype afternoon: three agents inserted 38 of 60 drafts, and replies fell from 3 minutes to about 2.6. Sixty tickets are enough to reject a branch, not to promise a rate, and stating the size lets the decider weigh it.
The bet, as a mechanism with its number. Drafts take a reply from 3 minutes to 2. At the measured 0.42 minutes saved, the 14 agents get back about 237 minutes a day, about 92 replies at 2.6 minutes against the 130 in the 4 pm backlog; at the full minute, 560 minutes, 280 replies at 2. The one-pager says both, because the gap between them is what the PRD has to close.
The cost, monthly and against the budget. $54,000 a month in tokens as specified, $27,500 after the two levers in tokens as a cost model, against a $30,000 budget and $60,000 of add-on revenue. Plus four engineers for a quarter, 12 person-months.
The line that stops it. More than 5% of sent replies wrong, the anxiety ceiling from the job, or the golden set under 92%. A one-pager without a stop line asks for an open-ended commitment, and a careful decider refuses those.
The ask: a decision, a person, a date. "Approve 12 person-months and a $30,000 monthly token budget; head of product; at Friday's planning review." Some teams put it first; either way it is on page one.
The PRD, section by section
The PRD keeps the one-pager's numbers and adds the ones a team needs to build. Eight sections, and the number each carries:
- Problem and outcome. The one-pager's count, verbatim, and the outcome metric with a baseline and a target: tickets carried overnight, 130 to under 20. Under it, the input the feature moves: minutes per reply, 3 to 2, measured on 60 timed replies per pilot tenant.
- Users and the job. The job statement with its count: clear the 130 by 6 pm.
- Scope and out of scope. Each exclusion with the number that put it there. No auto-send: at 8% of drafts wrong, sending them would put 16,000 bad replies a day in front of customers.
- Requirements. Functional ones as verbs; non-functional ones as numbers. "Fast" is not a requirement and "p95 under 4 s to the last token" is, the same rule the system-design course applies to capacity.
- Acceptance criteria. Each one a test with a pass condition. For a model feature this section is an eval, which the AI-feature PRD works through.
- Metrics and guardrails. The input the feature moves, and the three guardrails from AARRR and guardrails, each with a threshold and an action: p95 under 6 s per tenant per hour, complaints under 2 per 1,000 replies, cost per untouched resolution under $0.03.
- Launch plan. The stages with their exit condition: 5%, 25%, 100%, each held for a week of clean guardrails.
- Open questions. Each with an owner and the date it closes. An open question with no owner is a requirement nobody wrote.
The PRD's own version of the one-pager's mistake is a requirements section in adjectives. "Drafts should be accurate, fast and affordable" gets three builds: one engineer reads "fast" as 1 s, another as 10 s, and both are right about the sentence.
Where the two documents meet
The rule that follows from copying the one-pager's numbers is the most useful one here: when a PRD number drifts past a one-pager number, the decision is reopened, not the PRD quietly edited.
Say the eval clears the small model tier on half the tickets rather than 70%. On the same arithmetic as the token bill, the input line becomes $900 × (0.5 + 0.5 × 0.1) = $495 a day, output stays at $585 after the cap, and the month is $1,080 × 30 = $32,400, above the $30,000 the decider approved. The PRD does not absorb that. The PM sends the decider two lines, the new number and the option that brings it back (the cached prefix, once its rate has been read), and the decider decides again. A one-pager is a contract with one person, and the PRD is where its terms are kept.
In an interview
The questions come as "walk me through a spec you wrote" or "how would you get buy-in for this". The interviewer is testing whether you know who reads each document and what each section carries. Name the reader and the output, then give one number per section: "130 tickets carried overnight; 38 of 60 drafts inserted; $27,500 a month after two levers; approve 12 person-months by Friday."
The mistake that loses points is describing a document by its headings — problem, solution, requirements, timeline — with no number under any of them. The second is a one-pager with no ask: when the interviewer says "and what did you need from your manager?", the answer is a decision, a person and a date, not "alignment".
Check yourself
1. The head of product has five minutes before planning. Your one-pager runs to two pages, with a page of market background first and the ask at the bottom of page two. What do you cut, and what moves?
Cut the market background; no number in it bears on the decision. Move the ask to the first line: "approve 12 person-months and $30,000 a month, Friday". If it is still over a page, the evidence is carrying quotes; keep the counts — 40 interviews, 38 of 60 inserted — and link the snapshot for the quotes.
2. The bet assumes 1 minute saved per reply; the prototype measured 0.42. Which number goes in the one-pager?
Both, labelled: measured 0.42 on 60 tickets, target 1.0, and the discard rate by ticket type says where the gap is. At 0.42 the 14 agents get back about 237 minutes a day, about 92 replies at 2.6 minutes against the 130 in the backlog, a weaker case that clears about 70% of the queue, and the decider can weigh it. Sending 1 minute alone is the one error that cannot be corrected later without the decider discounting every other number.
3. Mid-build, the eval clears the small tier on 60% of tickets rather than 70%. Is that a PRD edit or a reopened decision?
Input becomes $900 × (0.4 + 0.6 × 0.1) = $414 a day; with output at $585 the day is $999 and the month about $29,970, $30 under the approved $30,000. A PRD edit, with one line to the decider saying the headroom is gone. The next drift of any size reopens the decision, so the PM should say now which lever is next.