/ ← → navigate N notes

Presenter Notes

Slide

⛰️ Contoso Outdoor Gear

Beyond Specs:
Agentic PM Workflows

How PMs use AI coding tools for knowledge management, prototyping, and workflow automation

Live demo with VS Code + GitHub Copilot • ~20 minutes
The Shift

PM tools are evolving

📝
2010s
Spec Writer
Jira, Confluence, Google Docs
🎬
2020s
Async Communicator
Notion, Linear, Loom
🧠
2025+
Prototype Contributor
VS Code, GitHub Copilot, GitHub Spark

"AI agents are a much better product manager than I ever was"

— Dennis Yang, Principal PM at Chime

Core Concept

The Product Brain

A markdown repo where AI can navigate, query, and synthesize all your product knowledge.

📁 Repository Structure
product-brain/
├── specs/           PRDs, feature specs
├── insights/        Synthesized research
├── research/        Market analysis, data
├── roadmap/         OKRs, priorities
├── analysis/        Jupyter notebooks
└── prototypes/      Working demos
🔍
Semantic Search
Query across all product knowledge instantly
🔗
Cross-References
Insights link to specs, specs link to roadmap
📜
Version History
See how decisions evolved over time
Building Blocks

Agent customization

📋 Workspace Instructions

Always-on context about company, conventions, personas

.github/copilot-instructions.md
📁 File Instructions

Auto-attach when editing specific file patterns

applyTo: "specs/**"
/ Prompts

Reusable workflows invoked with slash commands

/feedback-synthesis
🤖 Custom Agents

Specialized personas with tool restrictions

[Scaffold] from agent picker
Agents are selected via agent dropdown in VS Code, not @-mentioned
Live Demo

The Sustainability Dashboard

A real product initiative at Contoso. We have the spec and research — let's build it together, live.

1
Explore
What does Contoso know about sustainability?
2
Analyze
Dig into customer attitudes with notebooks
3
Prototype
Build the Personal Impact Dashboard
4
Challenge
Stress-test with Design Partner
🔍
Step 1

Explore the Knowledge

Before building anything, understand what Contoso already knows about sustainability.

Agent
Explore
Read-only research across the product brain
Ask
"What do we know about customer sustainability preferences?"
What it finds:
  • → specs/sustainability-dashboard.md — Approved spec with Impact Dashboard concept
  • → insights/sustainability-survey-2025.md — 68% say sustainability influences purchases
  • → research/gen-z-outdoor-trends.md — Gen Z expects eco transparency
  • → specs/loyalty-program-v2.md — Sustainability perks in Trail Rewards
📊
Step 2

Analyze Customer Data

The Explore agent found the sustainability survey insight. Let's dig deeper into the raw data.

Agent
Data Analyst
Creates Jupyter notebooks with EDA, stats, visualizations
Ask
"Analyze research/data/customer-survey-2025.csv — what predicts sustainability interest?"
What it produces:
📓
Jupyter Notebook
analysis/notebooks/sustainability-analysis.ipynb
📈
Cross-tab visualizations
Age × sustainability interest, Tenure × willingness to pay premium
🎯
Key finding
"Eco-Advocates" segment: high NPS, 2.3x LTV, 31% of base
⚡
Step 3

Build the Prototype

Now that we understand the opportunity, let's build a working demo of the Personal Impact Dashboard.

Agent
Scaffold
Builds HTML/Tailwind/Alpine.js prototypes
Ask
"Build a Personal Impact Dashboard prototype based on specs/sustainability-dashboard.md"
Spec highlights it will use:
  • • Lifetime CO2 offset counter
  • • Eco-certified purchase percentage
  • • Trees planted equivalent
  • • Shareable "Impact Card" for social
  • • Carbon offset calculator
Output: prototypes/sustainability-dashboard/index.html
🤔
Step 4

Challenge Our Thinking

Before stakeholder review, stress-test the approach. Find the blind spots before execs do.

Agent
Design Partner
Devil's advocate, Socratic questioning
Ask
"What's the strongest argument against specs/sustainability-dashboard.md?"
Devil's Advocate Mode

"How do you calculate CO2 offset without supplier data? Using industry estimates risks greenwashing accusations."

Socratic Mode

"What would need to be true for customers to actually share their Impact Cards? Have you validated that assumption?"

Blind Spot Hunter

"Who loses if this succeeds? Non-eco products could see cannibalization. Have you modeled that?"

The Pattern

What just happened

🔍
Explore
→
📊
Analyze
→
⚡
Prototype
→
🤔
Challenge
~15 min
From question to working prototype
0 meetings
No handoffs, no waiting for engineering
Full context
Citations, cross-refs, traceable decisions
More Workflows

What else can you do?

/ feedback-synthesis

Raw feedback → themed insight document with quotes

Paste 50 NPS comments, get structured analysis
/ spec-explainer

PRD → polished stakeholder page

Execs don't read 15-page specs — they scan beautiful summaries
/ meeting-prep

Gather context, surface open questions

"Prepare me for the checkout review with engineering"
/ issue-from-insight

Insight → GitHub Issue with labels & context

Traceable from customer quote to shipped feature
🤖 Diverge Agent

Explore 3+ variations via cloud coding agent PRs

"Compare minimal-mvp vs full-featured vs progressive-enhancement" — get actual PR branches to review
⚠️ Cautions

What to watch out for

"Almost right" is risky

AI-generated code and analysis needs human review. The 90% that looks correct makes the 10% that's wrong harder to spot.

Tool adoption ≠ Impact

Measure outcomes (conversion, customer satisfaction) not activity (prompts run, prototypes built). Easy to confuse motion with progress.

Role clarity matters

When PMs ship prototypes, who owns quality? When AI drafts specs, who validates claims? Define governance before scaling.

"Vibe coding your way to a production codebase is clearly risky."

— Simon Willison

Get Started

Try it yourself 

git clone https://github.com/digitarald/product-brain
code product-brain
1. Explore
Open any spec, ask about it
2. Try prompts
Type / to see available
3. Try agents
Switch via agent dropdown

7 demo scenarios in README • 7 custom agents • Real Jupyter notebooks • Interactive prototypes