Product Shaping Workflow

Ramp's Claude Code PM skill that turns customer signals into product specs — the operational heart of their VoC system.

The Three Phases

Phase 1: Frame the Problem

Claude asks the PM foundational questions:

• What is the job to be done?
• Why now? What changed?
• What does success look like?
• Who is the target user?

This constraints the research space before agents are launched.

Phase 2: Research with Parallel Agents

Claude launches 6-10 parallel agents simultaneously:

Agent	Data Source	Output
Gong Agent	Call transcripts	Customer quotes, pain points, frequency
Zendesk Agent	Support tickets	Top issues, workarounds, volume trends
Competitor Agent	Competitor websites/products	Feature gaps, positioning
Codebase Agent	Internal code	Technical constraints, existing patterns
Usage Agent	Product analytics	Adoption rates, drop-off points
CRM Agent	Salesforce	Deal context, segment patterns

Each agent writes a markdown artifact with structured findings. The coordinator agent then:

1. Reads all artifacts
2. Identifies cross-source patterns (themes appearing in 3+ sources)
3. Synthesizes into key findings with supporting evidence

Phase 3: Shape the Spec

Claude converges everything into a 2-minute-read spec containing:

• Context: Problem framing and research summary
• Design principles: Constraints and non-goals
• Requirements: Evidence-grounded, linked to specific customer quotes/data
• Alternatives considered: With trade-off analysis
• Open questions: What still needs human judgment

Why This Works

Speed

• 8 days of manual PM research -> 8 minutes
• Parallel execution means agents don't bottleneck on each other

Evidence Quality

• Every requirement links back to specific customer signals
• No "I think customers want X" — it's "47 Gong calls mention X, 23 support tickets describe the same pain"

Consistency

• Same process every time — no PM-dependent variation in research depth
• Junior PMs get senior-level research quality

Scale

• Ramp shipped 500+ features in 2025 with only 25 PMs
• The skill is the force multiplier — not headcount

Implementation Notes

Building Your Own PM Skill

1. Start with the framing questions — these constrain agent scope and prevent unfocused research
2. Define agent responsibilities clearly — each agent should have a narrow, well-defined task
3. Standardize artifact format — use consistent markdown templates so the synthesis agent can reliably parse
4. Tune for your data sources — replace Ramp's specific tools with your own (e.g., Intercom instead of Zendesk)
5. Add a human review step — the spec is a draft, not a final product

Artifact Template (Example)

## [Agent Name] Findings

### Top Themes
1. [Theme] — [X mentions, Y sources]
2. [Theme] — [X mentions, Y sources]

### Key Quotes
- "[exact quote]" — [Customer, Date, Source]

### Data Points
- [Metric]: [Value] ([Source])

### Confidence
- High/Medium/Low based on signal volume and consistency

Feedback Loops

Ramp maintains continuous improvement:

• Weekly reviews: Product ops partners read transcripts and compare to agent output
• Knowledge base updates: Corrections feed back into agent prompts and retrieval
• Sierra AI integration: Customer support agent has its own feedback loop — flagged issues surface to product

Key Takeaways

• The three-phase workflow (Frame -> Research -> Shape) is the core pattern — framing first prevents unfocused agent execution
• Parallel agents with standardized markdown artifacts enable map-reduce synthesis at scale
• Every requirement in the spec must link to evidence — this is what makes the output trustworthy
• The skill is a force multiplier: 500+ features with 25 PMs is the proof point
• Build continuous feedback loops — agent output quality degrades without ongoing calibration

Related

• Ramp VoC Overview
• Architecture Patterns
• Data Sources & Feedback Taxonomy