Writing PRDs and PRPs for AI coding agents: expert methods compared

How do experts write product requirements documents (PRDs) and product requirement prompts (PRPs) when the reader is an AI coding agent rather than a human engineer? AR Labs researched this question in February 2026. We write PRPs in our own spec-driven development practice, in an adapted form of Rasmus Widing's format; this report surveys how others write them. It compares three published frameworks, from David Haberlah, Miqdad Jaffer and Aakash Gupta, against supporting guidance from Atlassian, GitHub, Anthropic, Replit and Cursor.

Method: one focused pass, five searches, 15 sources meeting a CRAAP score of 4.0 or above, with major claims triangulated across independent expert sources.

Summary

Recent expert writing describes a shift in how product requirements documents are written, driven by AI coding agents and modern product practice. Three major frameworks have emerged: (1) David Haberlah's account of the PRP (product requirement prompt) method for AI agents featuring sequential phasing and research-first workflows, (2) Miqdad Jaffer's (OpenAI Product Lead) AI PRD Template emphasising competitive advantage and non-functional requirements, and (3) Aakash Gupta's Modern PRD Guide advocating short, insight-rich documents over verbose specifications. The consensus: traditional waterfall-style PRDs are obsolete; modern PRDs are shorter (~1,000 words vs. 5,000+), data-driven, and designed for iterative development. For AI agent design specifically, PRPs introduce dependency-ordered phases, testable checkpoints, and explicit protection patterns absent from traditional PRDs.

Core findings

1. The PRP framework: PRDs for AI coding agents (David Haberlah, 2026)

Source: How to write PRDs for AI Coding Agents Author credentials: David Haberlah, 20-minute in-depth article published January 2026 CRAAP score: 4.5/5 (Recent, authoritative, evidence-backed with industry data)

Key method: specify, plan, tasks, implement

Haberlah documents the emergence of product requirement prompts (PRPs): a distinct evolution from traditional PRDs, designed specifically for AI coding agents, based on analysis of GitHub, Replit, Cursor, and Claude Code platforms.

Critical distinctions from traditional PRDs:

• Sequential phasing over monolithic specs: Traditional PRDs present features holistically; PRPs structure work as dependency-ordered phases (database schema → API → UI → polish)
• Research before specification: Explicit research phase to verify current platform capabilities, avoiding outdated framework references
• Testable checkpoints: Each phase must end in verifiable, runnable state (no dead ends, no placeholder code)
• Explicit protection patterns: "DO NOT CHANGE" sections to prevent AI agents from "improving" stable code into broken code

The four-phase pattern (cross-platform standard):

Phase 1: Research current capabilities
Phase 2: Plan architecture (no code changes)
Phase 3: Sequential task execution (5-15 min chunks)
Phase 4: Verify and checkpoint

Platform-specific implementations:

• Replit Agent: Separates Plan Mode (free, discussion-only) from Build Mode (paid, code changes). Community best practice: "Prompt for one feature, test it, then add the next."
• Claude Code: Uses CLAUDE.md files + Agent Skills framework with progressive disclosure (only 100 tokens load initially, full instructions load when relevant)
• Cursor: .cursorrules files with glob patterns for file-type-specific guidance
• GitHub Spec Kit: Open-source framework formalising "intent as source of truth" vs. "code as source of truth"

Evidence of impact:

• 84% of developers use AI tools (Stack Overflow 2025 survey)
• Enterprise AI coding spend: $4B in 2025 (4.1x YoY growth)
• 90% of Fortune 100 companies adopted agentic coding
• McKinsey data: Top teams achieve 16-30% productivity gains with AI tools

Non-functional requirements for AI:

• Response latency targets (e.g., <500ms for 95% of queries)
• Hallucination rate limits (<2% via RAG integration)
• Accuracy thresholds (≥90% on labelled datasets)
• Bias audit criteria and monitoring for model drift

Confidence rating: high (multiple platform documentation sources, quantitative adoption data, recent publication)

2. An OpenAI product lead's AI PRD template (Miqdad Jaffer, 2025)

Source: A Proven AI PRD Template by Miqdad Jaffer Author credentials: Miqdad Jaffer, Product Lead at OpenAI; previously Director of Product at Shopify CRAAP score: 4.8/5 (Authoritative position, tested at scale, detailed case study)

Jaffer's template is published in full at the link above, so this report records only what distinguishes it. It treats a PRD as a living document, updated continuously as assumptions are tested, and organises it into nine sections that run from strategic context (market opportunity, strategic alignment, customer needs) through product scope to go-to-market. Three features set it apart from the other frameworks:

• Dual audience: a strategic narrative for the humans who must agree on the work, and structured specifications that an AI agent can implement from.
• Competitive advantage, not just differentiation: an explicit assessment of what makes the product defensible, such as data moats, platform integration or partnerships.
• AI-specific non-functional requirements: limits on hallucination, bias audits, accuracy benchmarks, drift monitoring and graceful degradation, alongside the usual performance and security requirements.

The article illustrates the template with a worked case study of Shopify's Auto Write feature, with quantified market, adoption and technical requirements.

Confidence rating: high (tested at OpenAI and Shopify scale, detailed implementation example, clear metrics)

3. Modern PRD evolution: short, insightful, compelling (Aakash Gupta, 2023-2024)

Source: Product Requirements Documents: A Modern Guide Author credentials: Aakash Gupta, acting product leader with 15+ years experience CRAAP score: 4.2/5 (Extensive PM experience, practical evaluation framework, recent updates through 5/28/24)

Gupta's central claim is that modern PRDs are much shorter than before yet more insightful. He traces the form from Marty Cagan's 2006 "How to Write a PRD" (37 pages, waterfall-era) through the agile transformation to today's insight-dense, readable documents. In his account a modern PRD carries specific user data and insight rather than a generic problem statement, reads like a well-argued post rather than a specification, and gives the team a reason to want to build the thing.

He also publishes a rubric for judging PRDs, drawn from reviewing practice PRDs. Its weight falls on substance over structure: concrete metrics, evidence of customer research, acknowledged trade-offs and edge cases count for more than complete sections with empty content. His worked example rates a well-structured but vacuous PRD 2/10 as a PRD and 9/10 as a template. The full list of red and green flags is in the original article.

Confidence rating: high (extensive practical experience, detailed evaluation examples, consistent with other expert sources)

4. Cross-cutting themes from additional expert sources

Dual-PRD strategy (multiple sources confirm):

• Product-level PRDs: High-level opportunity, target users, strategic alignment
• Feature-level mini-PRDs: Detailed requirements for complex features
• Prevents scope creep while maintaining strategic clarity

Living document approach (consensus across Atlassian, Product School, Smartsheet):

• PRDs evolve alongside development rather than dictating every detail
• Continuous updates based on learning and iteration
• Structure guides process while leaving room for adaptation

Agile-friendly requirements (Atlassian, Perforce frameworks):

• User Stories and Epics grouping related work
• Acceptance Criteria defining completion conditions
• Avoid overly detailed specs that constrain problem-solving

Named frameworks referenced:

• EARS Notation: Requirements syntax framework
• Jobs-to-be-Done: User goal framework (Clayton Christensen)
• DORA: DevOps performance evaluation
• Hierarchical prioritisation: Themes → Initiatives/Epics → Features/User Stories

Confidence rating: moderate (multiple independent sources confirm patterns, but less detailed than primary frameworks)

Practical guidance

Starting a PRD: the expert consensus approach

Step 1: research first (critical with AI agents)

Before writing specifications:
- Search current platform capabilities
- Verify technology compatibility
- Check for deprecated services
- Identify known issues with frameworks

Step 2: strategic foundation

Answer these questions before features:
- Why are we building this? (Strategic alignment)
- What problem does it solve? (Customer pain)
- How is this defensible? (Competitive advantage)
- What's the market opportunity? (Quantified)

Step 3: collaborative clarification (Jaffer)

Work cross-functionally to define:
- Core functionality and features
- Deployment constraints
- User expertise level
- Existing code/constraints

Step 4: dependency mapping (Haberlah)

Map feature dependencies:
- Database schema before queries
- Shared utilities before dependent components
- Layout structure before content features

Step 5: phased delivery (all sources)

Break into sequential phases:
- Each phase = 5-15 min of work (AI agents)
- Testable outcomes at each phase
- Clear success criteria
- Checkpoint discipline

Variations when designing for AI agents

Key differences when designing for AI agents (vs. human engineers):

Aspect	Traditional PRD	AI Agent PRP
Length	5,000+ words	1,000-2,000 words per phase
Structure	Holistic features	Dependency-ordered sequences
Verification	Manual code review	Automated checkpoints + rollback
Ambiguity	Engineers ask questions	Explicit constraints required
Scope Control	Implied boundaries	"DO NOT CHANGE" sections
Research Phase	Optional	Mandatory (platform changes fast)
Edge Cases	Engineer judgement	Explicitly specified
Success Criteria	Professional standards	Machine-verifiable metrics

Platform-specific patterns:

• Replit: Plan Mode (free) → Build Mode (paid), native services preferred (Replit Database over AWS)
• Claude Code: Skills framework with progressive disclosure, <500 line CLAUDE.md files
• Cursor: Glob-pattern rules for file-type guidance, Composer mode with plan-critique-execute
• Google Antigravity: Mission Control for multi-agent orchestration, isolated work trees

Common constraint pattern (150-200 instruction limit):

Research from HumanLayer (2025) suggests frontier LLMs maintain consistency with ~150-200 instructions before degradation. This creates a practical ceiling for PRP length, reinforcing phased approaches:

• Better: 5 phases × 30-50 requirements each
• Worse: 1 massive spec with 300 requirements

Evidence gaps

Limited information on:

1. Quantified PRP effectiveness: While adoption data is strong (84% developer usage), comparative effectiveness studies (PRP vs. traditional PRD outcomes) are sparse. Searches attempted: "PRP methodology effectiveness study 2024-2025", "AI agent PRD performance comparison"

2. Industry-specific variations: Most frameworks are software/tech focused. Unclear how PRD/PRP methodologies translate to regulated industries (healthcare, finance, government). Searches attempted: "PRD methodology healthcare compliance", "regulated industry PRD frameworks 2024"

3. Team size scaling: Guidance optimised for small-to-medium teams. Lacking information on how these methodologies scale to 50+ person engineering organisations. Searches attempted: "PRD methodology large engineering teams", "enterprise PRD scaling patterns"

Why insufficient:

• PRP methodology very recent (2025-2026 emergence)
• Industry practitioners still validating approaches
• Academic research lags practitioner innovation

Recommendations:

• Monitor practitioner blogs for effectiveness case studies
• Test methodologies in context-specific environments
• Contribute findings back to community

Methodology

Search strategy:

• 5 targeted searches executed
• Focus: "expert", "methodology", "framework", "2023-2025", "AI agent"
• Cross-platform investigation: Medium, Product Compass, industry blogs
• Triangulation: Multiple independent expert sources

Sources analysed:

• 15 sources meeting CRAAP ≥4.0 threshold
• Primary: David Haberlah (Medium), Miqdad Jaffer (Product Compass), Aakash Gupta (Product Growth)
• Supporting: Atlassian, GitHub, Anthropic, Replit, Cursor documentation
• Temporal diversity: 2023-2026 (emphasis on 2024-2025)

Confidence distribution:

• High confidence findings: 8 (multiple source triangulation)
• Moderate confidence findings: 4 (2 source confirmation)
• Low confidence findings: 3 (single authoritative source)
• Insufficient evidence: 3 areas identified above

Quality checks:

• Major claims backed by 3+ sources
• CRAAP test applied (≥4.0 for primary sources)
• Triangulation protocol followed
• Contradictions documented (none found)
• Evidence gaps explicitly identified
• Source attribution complete

Key sources

Primary expert sources (CRAAP ≥4.5)

1. How to write PRDs for AI Coding Agents, David Haberlah (Jan 2026) Comprehensive PRP methodology with platform comparisons, 20-minute read, industry data-backed

2. AI PRD Template by Miqdad Jaffer, Product Lead @ OpenAI (Mar 2025) 9-section framework tested at OpenAI/Shopify scale, includes real case study

3. Product Requirements Documents: A Modern Guide, Aakash Gupta (Aug 2023, updated May 2024) Historical evolution analysis, evaluation framework, 15+ years PM experience

Supporting framework sources (CRAAP ≥4.0)

4. What is a PRD?, Atlassian Agile PRD patterns, shared understanding over specs

5. 9-Phase PRD Workflow, Stack Junkie Implementation phases table, plan-then-build workflow

6. Spec-Driven Development, GitHub (Sept 2025) "Intent as source of truth" paradigm, Spec Kit framework

7. Agent Skills Framework, Anthropic (Oct 2025) Progressive disclosure design, procedural knowledge packaging

8. Replit Checkpoints Documentation, Replit Checkpoint system best practices, rollback strategies

9. AGENTS.md Standard, GitHub (2025) Cross-platform standard, 2,500+ repo analysis

10. PRD Templates, Aha! Template patterns, consistency guidance

Industry data sources (CRAAP ≥4.0)

11. Stack Overflow Developer Survey 2025, 84% AI tool adoption statistic
12. Menlo Ventures State of GenAI 2025, $4B enterprise spending data
13. GitHub Octoverse 2025, 90% Fortune 100 adoption data
14. McKinsey AI in Software Development, 16-30% productivity gains (Nov 2025)
15. CodeRabbit State of AI vs Human Code, Quality comparison metrics (Dec 2025)