2.1 Write a PRD, keep the thinking

An agent will write you a complete PRD in ninety seconds. It will be well-organised, correctly formatted, and hollow — because the hard part of a PRD is not the prose, it is deciding what not to build and being able to say why.

This method uses the agent at the six points where it actually helps.

Phase 1 — Choose the shape

Decide the format before there is any content, or you will retrofit your thinking into whatever the agent produced. Two that work:

Two PRD shapes
Shape	Sections	Use when
Full	Problem alignment, solution alignment, development plan, open questions	Cross-team work, or anything a director will read
Minimal	What, why, who, how, success measure, open questions, timing	One team, known problem, weeks not quarters

Save the one you pick as context/prd-template.md. It becomes a rule rather than a preference.

Phase 2 — Build the context

Attach explicitly. Never rely on the folder being open.

I am writing a PRD on onboarding activation for TaskFlow. Context: @context/company.md, @context/rules.md, @research/interviews-march.md, @research/support-export.csv, @shipped/last-prd.md for house style. Do not write anything yet.

The last clause matters more than it looks. Without it you get a draft immediately, and from then on you are editing rather than thinking.

Phase 3 — Be interrogated

Ask me the eight questions a sceptical director would ask before approving this, one at a time. Do not answer them yourself. Start with the one I am most likely to have avoided.

Answer them in writing. This is the phase people skip and it is the phase that produces the document. Your answers, lightly edited, become the problem-alignment section — and unlike generated prose, they contain things only you know.

Phase 4 — Get options, not an answer

Based on my answers and the attached research, draft three genuinely different solution directions. They must differ in what they refuse to do, not in how they are phrased. For each: who it serves best, who it fails, and what it costs us to be wrong.

“Differ in what they refuse to do” is the working part of that sentence. Without it you get one idea in three registers.

Phase 5 — Review from three hostile angles

Run these separately, not as one request. Combined, they average into mush.

The engineer. “Read as Mike. What is underspecified to the point where you would have to guess? What edge case will bite us in week two?”
The admin buyer. “Read as Sarah. What here is impossible to govern, audit or roll out to 400 seats?”
The lead. “Read as Alex. If this ships exactly as written, what can you still not see about your team’s work?”

Phase 6 — Decide, then record why

Address the feedback yourself. Then finish with the section that makes a PRD survive contact with next quarter:

Write an “alternatives considered” section: the three directions, which was chosen, and the specific reason each other one was not. Use my stated reasons. Do not invent a rationale I did not give.

Where this goes wrong

Phase 3 is skippable and skipping it is fatal. A PRD built from phases 1, 2, 4 and 6 reads beautifully and contains no proprietary judgement, which is exactly what a director notices and cannot articulate. If you only have twenty minutes, do phase 3 and write the rest by hand.