PM Agent — AI Product Management Workflow

Four agents covering the full PM lifecycle: Research → Define → Validate → Launch. Each phase uses proven frameworks and produces structured artifacts. Human checkpoints between phases.

Phases

How to Use

Full Workflow

"I want to build [product idea]" → run all 4 phases
"Run pm-agent on [problem statement]"

Each phase spawns a focused subagent with the right prompt. The subagent asks questions, challenges assumptions, and produces a structured artifact.

Partial Workflow

• "Just do a competitor analysis" → run Research only
• "Help me prioritize my backlog" → run Define (RICE/Kano section)
• "Write a PRD for this feature" → run Define with the feature description
• "Plan a design sprint" → run Validate only
• "Create a GTM plan" → run Launch only

Single Commands

• /research — JTBD interview analysis, market sizing, competitive landscape
• /define — Opportunity Solution Tree, PRD with Amazon Working Backwards
• /validate — Experiment design, prototype testing plan, BML metrics
• /launch — GTM strategy, OKRs, release checklist

Phase Details

Phase 1: Research (JTBD + Design Thinking)

Goal: Understand the problem before proposing solutions.

Spawn a subagent (Sonnet) with the Research prompt from references/prompts.md. It will:

1. JTBD Analysis — Extract the "job" users are hiring the product for
• Push factors (pain with current solution)
• Pull factors (attraction of new solution)
• Trigger event (what moment starts the search)
2. Competitive Landscape — Map existing solutions and gaps
3. Market Sizing — TAM/SAM/SOM with reasoning
4. User Personas — 2-3 evidence-based personas (not fictional)
5. Write DISCOVERY.md — Consolidated research artifact

Key question: "What job is the user hiring this product to do?"

Phase 2: Define (Opp. Tree + Amazon PRD)

Goal: Define what to build and why, before how.

Spawn a subagent (Sonnet) with the Define prompt. It reads DISCOVERY.md and produces:

1. Opportunity Solution Tree — Visual hierarchy of outcome → opportunities → solutions
2. Prioritization — RICE scoring for top opportunities, Kano classification
3. Amazon PRD — Working Backwards: start with the press release, then FAQ
4. User Stories — INVEST-compliant stories with acceptance criteria
5. Write PRD.md — Complete product requirements document

Key rule: No solution before opportunity. No feature before user story.

Phase 3: Validate (Design Sprint + Lean)

Goal: Test assumptions before building.

Spawn a subagent (Sonnet) with the Validate prompt. It reads PRD.md and produces:

1. Assumption Map — Classify by risk (lethality × uncertainty)
2. Experiment Design — Lean BML cycle for riskiest assumptions
3. Prototype Plan — What to mock up and how to test with 5 users
4. Success Metrics — Quantitative pass/fail criteria per experiment
5. Write EXPERIMENT.md — Validation plan with test scripts

Key rule: Test the riskiest assumption first, not the easiest.

Phase 4: Launch (GTM + OKR)

Goal: Ship and measure.

Spawn a subagent (Haiku) with the Launch prompt. It reads PRD.md and EXPERIMENT.md and produces:

1. GTM Strategy — ICP, positioning, channel mix
2. OKRs — 3 measurable objectives with key results
3. Release Checklist — Pre-launch, launch day, post-launch tasks
4. Feedback Loop — How to collect and act on user signals
5. Write GTM.md — Launch plan with timelines

Key rule: Launch is not the end. It's the beginning of the BML cycle.

Model Selection

Output Files

All phase outputs go to the project root:

• DISCOVERY.md — Research findings (JTBD, personas, competitive landscape)
• PRD.md — Product requirements (Opp. Tree, Amazon PRD, user stories)
• EXPERIMENT.md — Validation plan (assumptions, experiments, metrics)
• GTM.md — Launch plan (GTM, OKRs, checklist)

Each file is self-contained but references previous phases. You can run phases independently by providing the prerequisite context.

Frameworks Reference

For detailed framework guides (JTBD interview templates, RICE calculators, Amazon PRD templates), see references/frameworks.md.

Human-in-the-Loop

Each phase ends with a checkpoint:

• Approve — proceed to next phase as-is
• Edit — modify the artifact, then proceed
• Rerun — provide feedback, regenerate the phase

This mirrors real PM work: AI drafts, humans decide.