Tasks / Define

Write a PRD

Can the model turn a brief into a spec engineers could build from, including how an AI feature behaves when it is wrong?

Measures the modelTask v1.1 · 2 casesDifficulty

The PM job

Writing the requirements document a team will build and test against.

Why it matters

A PRD is where ambiguity becomes either a decision or a bug. For AI products it must also say what happens when the model is uncertain or wrong. Most generated PRDs skip that part.

What good looks like

  • States the user problem and the decision the PRD enables
  • Specifies behaviour under uncertainty, failure and refusal
  • Names eval criteria and a launch bar
  • Separates must-haves from later ideas

Deliberately not measured

  • Formatting or template conformance
  • Length
  • Visual polish of diagrams
Capability tested

Making product behaviour, uncertainty and eval requirements executable

The failure we’re looking for

A generic feature spec that ignores AI failure behaviour

Grading

Decision model, LLM judge and blind PM review

Variants

AI product PRD (core) · Conventional product PRD

Results

Every setup we’ve tested on this task, across all cases and repeats.

#Model · HarnessTask scoreDecision modelLLM judgePM reviewRunsCritical failuresCost / runLatency

Case viewer

Read the brief, then put up to three outputs side by side. The outputs are the point; the scores just tell you where to look.

The brief

Write a PRD for an AI feature that drafts first responses to support tickets and routes each ticket to the right queue. Support agents approve every draft before it is sent. The PRD is for the engineering, ML and support-operations leads who will build and run the feature: write it so they could start work from it. Aim for 1,000–1,500 words.

ScenarioLedgerly is accounting software for small businesses. The support team is 42 agents across four queues: Billing, Technical, Account access and General. Two years of resolved tickets, with the agents' replies, are available for training and evaluation.
Volumes9,000 tickets a week; 38% are billing. Average first response is 7 hours against a target of under 2. About 60% of tickets are answered with one of roughly 40 saved replies.
RiskLegal requires no automated sending of refund commitments.
What a strong answer does

A PRD that defines behaviour when confidence is low, forbids refund commitments, sets an eval set and launch bar, and specifies human oversight.

Critical failures (cap the score)
  • Allows automated refund commitments
Case

v1.3 · synthetic · AI product, support