The best PRDs I've written don't just specify what to build. They document why we chose this over that, what we're explicitly not building in this version, and how we'll know it worked. If you're hiring a PM to write specs, these show how I think before engineering starts.
All PRDs here are sanitized where necessary to protect client-specific information. The structure, decision logic, and acceptance criteria represent real specification work.
Full PRDs
PRD 01
AI · Data Product · B2C
Multi-persona API
JobSignals — Labor Market Intelligence Platform
Full PRD for a zero-to-one AI data product that extracts structured hiring signals from unstructured job posting text. Six distinct user personas, REST + GraphQL API, and a prompt engineering specification that I wrote as acceptance criteria. The insight: writing precise prompts for consistent AI output taught me more about writing good feature specs than any methodology.
Signature Decision in This PRD
Chose coverage over accuracy. The salary signal extraction trades precision for breadth — extracting probabilistic salary ranges from implicit language ("competitive compensation" + seniority signals) rather than explicit numbers. Documented with explicit confidence interval design so users can calibrate their own trust. This tradeoff decision is the center of the PRD, not a footnote.
PRD Structure
Problem Statement
User Personas (6)
Non-Goals
API Specification
Prompt Engineering Spec
Acceptance Criteria
Success Metrics
Open Questions
Zero-to-one
Claude API
Data product
REST/GraphQL
Read PRD →
PRD 02
B2B Fintech · KYC
RBI Compliance
KYC Guided Capture — Feature PRD
Feature spec for the guided document capture experience that moved KYC first-attempt success rate by 28% in the first month at Signzy — before touching the AI model. Every acceptance criterion has a compliance annotation explaining why the constraint exists (not just what it is). Designed for two users simultaneously: the bank customer completing KYC, and the bank ops team reviewing the results.
Signature Decision in This PRD
Real-time quality feedback specified at the component level, not the feature level. Rather than "add document quality check," the spec defines the exact feedback states (lighting_low, edge_cropped, blur_detected, quality_ok), the latency threshold (feedback must appear within 200ms of frame stability), and the fallback when the quality check fails to load. Compliance annotation: "No quality score or image data may be retained on device after capture session ends — RBI data minimization requirement." Engineering never needs to ask whether to retain data.
PRD Structure
Problem Statement
North Star Metric
Users (End Customer + Ops Team)
Non-Goals
Feature Specification
Acceptance Criteria + Compliance Annotations
Edge Cases
Success Metrics
Open Questions
RBI compliance
Mobile-first
AI-assisted UX
B2B Fintech
Read PRD →
PRD 03
GenAI Feature
Healthcare · HIPAA
AI Email Drafting — TruServe Feature PRD
Feature spec for smart email drafting — the Phase 1 P0 feature from the HCSC $10M program that achieved 78% agent acceptance rate in Sprint 4. The most interesting section: the AI confidence score UX specification, designed explicitly for skeptical agents who need to feel in control of AI output, not replaced by it. Also includes the override design and how override data feeds back into model improvement.
Signature Decision in This PRD
The override UX is specified as carefully as the generation UX. Most AI feature PRDs specify what happens when the AI works. This one specifies what happens when agents don't like the output: one-click override, pre-populated edit mode (not blank slate), and an implicit feedback signal captured from every override (which phrases were edited, what replacements were made). The override isn't a failure state — it's a data collection mechanism that improves the model over time. This design decision is why adoption was 30% higher than our baseline assumption.
PRD Structure
Problem Statement
Risk Assessment (Quadrant D classification)
Users (Agent + Supervisor + IT)
Non-Goals
Generation Spec
Confidence Score UX
Override Design + Feedback Loop
HIPAA Compliance Notes
Acceptance Criteria
Adoption Metrics
GenAI
Adoption design
Healthcare SaaS
Enterprise
Read PRD →
PRD 04
Zero-to-One
B2B SaaS
Founder OS — Operating Layer PRD
Full PRD for a zero-to-one product with no brief, no existing product, and no validated demand. The challenge: there are 20 existing tools for different parts of what founders need. The PM question at the center of this PRD: what is the minimum slice of functionality that creates irreversible founder habit — the thing that makes a founder open this product every morning instead of switching between the six tools they already have?
Signature Decision in This PRD
Chose OKR tracking as the wedge feature — not investor CRM, not burn rate dashboard, not roadmap. The hypothesis: founders open their OKR tool at the start of every week. If we own the Monday morning review, we own the context that makes every other feature (investor updates, hiring decisions, sprint planning) naturally want to connect to us. The PRD documents this reasoning explicitly in the "why this feature is the wedge" section — because zero-to-one PRDs need to explain the sequencing logic, not just the feature list.
PRD Structure
Problem Statement
Market Context
User Research Synthesis
The Wedge Feature + Rationale
Module Specifications (6 modules)
Integration Architecture
Pricing Model
Non-Goals
Success Metrics
Zero-to-one
B2B SaaS
Founder tools
Pricing strategy
Read PRD →