Guide · Product management
How to write a PRD engineers can estimate.
Eight steps from a feature idea to a product requirements document your team can break down, estimate as a range and sign off, with a worked example.
In short. A product requirements document (PRD) says what a feature has to do and why, before anyone decides how to build it. A good one names the problem, the people it touches, what done means, what is out of scope and the open questions, so engineering can break it down and estimate it as a range.
What a PRD is for
A PRD is a shared answer to three questions: what problem are we solving, for whom, and how will we know we’ve solved it? It is owned by product and read by engineering, design and the owners of anything the feature touches, such as billing, security or support.
It is not a technical design. The PRD says what and why; the technical design says how. Keeping them apart lets people argue about the right thing: scope in the PRD, architecture in the design doc.
Most of a PRD’s value comes from the questions it forces before the sprint, not from the document itself. Each question answered on paper is one that doesn’t turn up halfway through the build. More on that in what causes scope creep.
For a small feature, a one-page feature description template is often enough.
Before you write
Collect three things first.
- The evidence. Support tickets, interview notes, usage numbers, a sales request. Link them; don’t paraphrase them.
- The people. Who starts the flow, who is on the other end, and who owns the areas it touches. If the feature sends email, whoever owns email deliverability is a reviewer.
- The constraint. Is there a date, a contract or a launch it has to fit? Say so up front, because it decides what goes in “Not now”.
How to write a PRD, step by step
Work through the sections in this order. It is not the order they appear in the finished document: the summary comes last, once the scope has settled. The PRD template has every section with prompts, and the PRD generator fills it in as you type.
- Write the problem, without the solution. Describe the situation today: who has the problem, how often, and what they do instead. Reviewers judge your solution against this paragraph, so if it already contains the answer, there is nothing to judge against. Weak: “We need an invite button.” Better: “Admins add teammates by sending them the sign-up link in chat. New people land in a workspace of their own and have to ask support to move them.”
- Name every user and role. List each kind of person the feature touches, including the ones who only receive something. Then write the permission rules: who may do it, and who must not. This is the part engineers most often have to guess, and a wrong guess is expensive to undo.
- Write “Done when” as checks a tester could run. Each line should be something you can verify on the real screen. “Improve onboarding” can’t be tested. “An admin can invite up to 20 people in one go” can. Add what you’ll measure afterwards and today’s number if you have one. If you don’t, say so rather than inventing a target.
- Write “Not now”, with a reason for each item. List what people will assume is included but isn’t: the obvious next feature, the edge case you’re accepting, the platform you’re not supporting yet. Anything not on this list will be asked for in review, and it is far easier to agree on now.
- Break the requirements down by layer. Group what has to exist by where it lives: screens, API, data, and the rules that cut across them, such as permissions, rate limits, logging and rollout. For each screen, list the states: empty, loading, error, no access. Mark each requirement must, should or could, and describe behaviour, not implementation. Nobody can estimate “team invitations”; people can estimate “invite email” or “accept-invite endpoint”. The guide to breaking down a feature goes step by step.
- Walk the edge cases and write down open questions. Each edge case becomes either a requirement or an open question, and every open question gets an owner and a date to decide by. The list below is where to start.
- Estimate as a range, not one number. Ask the engineers who will build it for best, likely and worst days per requirement. A single number is read as a promise; a range shows how far the bad case is from the likely one, and the parts with the widest spread tell you where to spike first. Turn the ranges into the dates you’d commit to at 50, 80 and 90% confidence. The free estimation tool does that from your three numbers per part. If your team sizes in points instead, see story points explained.
- Get sign-off, then write the summary. Send it to product, engineering, design and the owner of every area it touches, and record who approved and any condition they attached. Write the three-sentence summary last. If it won’t fit in three sentences, the scope hasn’t settled yet.
Edge cases to walk through
These are the cases that most often turn up mid-sprint when nobody asked about them in planning. An open question without an owner turns into scope creep.
- It already exists: the invited email already has an account, or is already a member.
- It expires or is used twice: a link clicked after it expired, or clicked twice.
- It isn’t allowed: someone without the right role, or a workspace at its seat limit.
- It fails: the email bounces, the payment is declined, the background job dies.
- Everyone uses it: keyboard and screen reader, slow networks, other languages.
A short example: team invitations
Here are four sections of a PRD for a feature where workspace admins invite colleagues by email. The full version is in the PRD template.
Users and roles. Owners and admins send invites and choose the role. The invited person receives an email and joins when they accept. Members can’t invite.
Done when. An admin can invite up to 20 email addresses at once and pick a role for each. Invite links expire after 7 days and work only for the address they were sent to. Admins can resend and revoke pending invites.
Not now. Invite links without an email address, because they get around the seat limit and the address check. SSO-only workspaces, because they need their own way to join.
Open question. Does the product charge per seat? If so, what happens when an invite would go over the limit? Owner: billing. Decide before estimating.
Notice the open question. It is the kind that rarely comes up in the planning meeting and then arrives in week two as “billing says invites have to respect seat limits.”
Common mistakes
Most weak PRDs fail in one of these ways.
- Writing the solution as the problem. “Users need a dashboard” is a solution. What can’t they see today, and what does that cost them?
- No “Not now” section. Without it, every reviewer adds the thing they assumed was included.
- “Done when” that can’t be tested. If two people could disagree on whether it’s done, rewrite it.
- Edge cases left to engineering. They’ll be handled, but without product deciding, and usually mid-sprint.
- One-number estimates. “Two weeks” becomes the date, and the worst case is never discussed.
- Writing it alone. A PRD nobody reviewed is a wish list. Ask the tech lead to read the first draft, not the final one.
- Too long. For a feature of a few weeks, one to three pages is usual. If a section has nothing to say, write “none” so readers know it was considered.
How to know it’s ready
A PRD is ready for the sprint when all of these are true. If one isn’t, the gap will show up during the build instead.
- An engineer who wasn’t in the meeting can break it into parts without asking what it means.
- Every open question has an owner and a date, or has been answered.
- Each part has a best, likely and worst estimate, and you can say which part is riskiest.
- The owners of every area it touches have signed off.
- You’d be comfortable sharing the 80% date with whoever is waiting for it.
Doing this in SprintProof
SprintProof takes a feature through these steps. Its AI interview asks who the feature is for, what done means and what is out of scope, one question at a time. The plan is then broken into parts with best, likely and worst days, checked against a library of edge-case questions, reviewed and signed off by the team, and estimated as 50, 80 and 90% dates.
The AI suggests and people accept: nothing in the plan changes without a click. See pricing for the Team plan.
Plan the feature, not just the document.
Describe the feature and SprintProof interviews you, drafts the breakdown, asks the questions reviewers ask, and gives an estimate as a range with 50, 80 and 90% dates.