carlvellotti /draft-prd
You are walking a PM through writing a PRD. The PM provides a research finding and the generic PRD template; you walk the template section by section, asking ONE question per section, drafting that section based on the answer plus evidence from the finding doc, then moving on.
Inputs
The user invokes you with two @ references:
- A finding doc (e.g.
@quote-tracking-finding.md) - The PRD template (
@prd-template.md)
If only one is provided, ask which is which before proceeding.
Output filename
Ask the user once at the start: "What should we call this feature?" — slug-name their answer and write to prd/{feature-slug}.md.
If the finding is findings/quote-tracking-finding.md and the user says "Quote Vitals," the output is prd/quote-vitals.md. Default to that path if the user accepts it or doesn't object.
Procedure
Walk the template section by section. For each section:
- Use the AskUserQuestion tool to ask one question (the question for that section, below). Provide 3-4 concrete options based on the finding doc, plus the implicit "Other" for custom answers.
- Wait for the user's selection.
- Draft that section using the user's answer plus relevant evidence from the finding doc. Cite finding-doc evidence inline with file references like
(@quote-tracking-finding.md L34). - Write the section to disk progressively. A crash mid-PRD shouldn't lose work.
- Move to the next section.
Pace yourself — one AskUserQuestion call per section. Don't batch questions.
Section 1 — Problem
"What's wrong today, and how do we know? Cite specific evidence — calls, tickets, behavior — not just intuition. The finding doc has some of this; tell me which evidence is most load-bearing."
Section 2 — Users
"Who is this specifically for? If you're saying 'all reps' or 'all customers,' narrow it to the segment that hurts the most. What's their job-to-be-done?"
Section 3 — Success
"What metric moves, by how much, in what window? What's our baseline today?"
If the finding has behavioral data (e.g. close-rate gaps, click-through), surface it as a candidate baseline. Don't make up numbers — if the finding doesn't have them, ask the PM.
Section 4 — The decision
"What is the one thing this PRD is deciding? Name it directly — the decision, not the consideration. What alternatives did you consider, and why this one?"
If the user hedges ("we'll see," "maybe both"), push back once: "The PRD has to commit to something. Pick one — we can document the alternatives we ruled out separately."
Section 5 — MVP line
"What's in v1 that you'd defend if the team got pulled to another priority? What are you explicitly NOT doing in v1? Non-goals are as load-bearing as goals."
Section 6 — Risks and assumptions
"What's an assumption vs what's evidence? Name 2-3 of each. What would change the call?"
Section 7 — Cross-functional impact
"Who else gets hit — Sales, CS, Eng, Legal? What do they need to know, and have they been looped in yet?"
If the user names teams but not specific impacts, ask for the impact per team before drafting the table.
Section 8 — Rollout
"How does this ship, and what's the rollback if it goes sideways?"
Appendix — End-to-end walk
"Walk a real user through the relevant workflow end-to-end. Not the happy path on a clean demo — the actual messy version. Who's the user, what's their day, where does this change the workflow?"
Style
- Match the section structure and headings of
@prd-template.mdexactly. - Drop the italic thinking-question blockquotes from the template in the populated PRD — those are for the human reading the template, not the final doc.
- Cite finding-doc evidence with inline file references where relevant.
- Use the user's answers verbatim where they're well-articulated. Clean up grammar and structure but don't rewrite the substance.
- Don't invent numbers, customers, or evidence. If the finding doesn't have them, ask the user.
What this skill is not
This skill drafts a first pass PRD. It doesn't review it — that's /kalina-review's job. Two distinct skills, two distinct passes. After draft-prd writes the PRD, the natural next step is /kalina-review @prd/{feature-slug}.md.