PM Approach - How I Think & Work
This page covers my approach to product management: how I think, how I prioritise, how I write specs, how I build with AI, and what I believe makes a great PM.
My Discovery Framework
I follow a structured discovery process before writing any spec:
1. Understand the business context
→ What metrics matter? What’s the OKR this connects to?
2. Map the current state
→ What does the user do today? Where do they struggle?
3. Generate hypotheses
→ Why does this problem exist? What are 3 possible causes?
4. Validate ruthlessly
→ Interviews, data, observation - in that priority
5. Frame the problem precisely
→ “Users can’t do X because of Y, which leads to Z business impact”
6. Explore solution space widely, converge carefully
→ 3+ approaches before picking one
How I Write PRDs
My PRD template always includes:
| Section | What I write |
|---|---|
| Problem statement | The specific user pain, validated with data/quotes |
| Why now | Why this over everything else we could build |
| Success metrics | Specific, measurable, time-bound. Not “improve conversion” - “increase meeting-to-applicant rate from 15% to 17% in 6 months” |
| Non-goals | Explicitly what we are NOT building. Usually as important as goals. |
| Design decisions | Key tradeoffs made, with rationale. I want engineers to understand why. |
| Open questions | What I don’t know yet. Tracked until resolved. |
| Risk log | What could go wrong, and how we’ll mitigate |
My Approach to AI Product Development
Building with LLMs and ML is different from building traditional software. My principles: 1. Start with the task, not the technology The question is never “can we use an LLM here?” - it’s “what does the user need to accomplish, and is an LLM the right tool?” 2. Explainability is a feature, not an afterthought For every AI product I’ve built, user trust depended on understanding why the AI made a decision. I build explainability into the design spec from day one. 3. Define the human-in-the-loop early AI systems make mistakes. Before shipping, I define: what mistakes are acceptable? At what confidence threshold do we route to a human? Who is that human and what’s their workflow? 4. The feedback loop is the product An AI product that doesn’t get smarter over time is just an expensive static feature. I design feedback collection into every AI product from v1. 5. Prompt engineering is product work The quality of an LLM product is determined by the quality of its prompts and schemas. I write and iterate on prompts with the same rigour as I write product specs.
How I Prioritise
| Dimension | Notes |
|---|---|
| Revenue / business impact | Direct revenue, churn prevention, upsell enablement |
| User impact | How many users? How severe is the pain? How often? |
| Strategic alignment | Does this build toward our 12-month vision? |
| Effort (inverse) | Lower effort scores higher |
I share this in Excel with stakeholders so prioritisation decisions are transparent and challengeable.
How I Run Discovery Interviews
My interview principles:
- Ask about behaviour, not opinions. “Tell me about the last time you did X” not “would you use feature Y?”
- Follow surprises. The most valuable insights are the ones you didn’t expect.
- Never pitch during discovery. I’m there to learn, not to sell.
- Synthesise in 24 hours. Patterns fade fast. My standard question sequence:
- Walk me through your day / a typical [task]
- Tell me about the last time you did [specific thing]
- What was the hardest part?
- What did you do to work around it?
- What would “good” look like for you?
- What else should I know?
How I Think About Metrics
I start with the growth equation for the product and identify the lever we need to move—acquisition, activation, conversion, retention, or monetisation. Before prioritising a feature, I model its expected impact: Eligible users × expected adoption × expected behaviour change × business value I then sensitise the assumptions across conservative, expected, and upside scenarios. This helps compare opportunities based on potential impact, confidence, and effort—not just stakeholder demand. Once shipped, I track:
- Leading success metrics: early behavioural signals that show whether the feature is working
- Output metrics: the direct customer or operational change the feature was designed to create
- Lagging metrics: the eventual business impact on conversion, retention, revenue, churn, or NPS For example: Recommendation feature → recommendation usage → higher course shortlisting → improved meeting-to-applicant conversion This connects every feature to the growth equation and makes success measurable before, during, and after launch.
What I’m Exploring Now
- Designing products around repeat behaviour — how progress visibility, continuity, personalisation, and timely interventions can improve retention. MyYogaTeacher — Repeat Session Retention
- Building sustainable growth loops in consumer marketplaces — connecting discovery, participation, community, supply, and rewards to increase repeat usage. Playo — Growth and Participation Loop
- Turning fragmented operations into decision systems — creating a shared intelligence layer across operations, finance, marketing, coaches, trainers, and leadership. Cult — Operations Intelligence Platform · Interactive Prototype
- AI-native customer journeys — exploring how recommendations, conversation intelligence, adaptive interfaces, and progressive disclosure can reduce decision friction and improve conversion.
- Agentic workflows in B2B products — designing automation loops that can plan, execute, evaluate outcomes, and improve with limited human intervention.
- Products that connect behavioural and health data to action — translating wearable signals, training patterns, and user goals into clear recommendations rather than static dashboards.