Product sense — recap & real-world examples

Real-world examples & war stories

Slack's behavioural segmentation. Rather than "team communication for everyone," Slack targeted engineering teams already chatting in IRC — a segment defined by behaviour, not demographics. It won a beachhead that then expanded. 🎯 Takeaway: the sharp strategic-thinking move is segmenting by what people already do. That's exactly the instinct an AI PM needs to find the job a model reliably nails.

Amazon's Kindle vision and the PRFAQ. "Every book ever printed, in any language, all available in under 60 seconds" is a vision that made thousands of decisions for the team. The PRFAQ format forced every proposal to start from the customer. 🎯 Takeaway: strong communication artifacts turn product sense into alignment. For an AI feature drowning in uncertainty, an explicit "what it is / isn't / how we'll know it worked" is worth more than any demo.

LinkedIn's profile-completeness bar. A progress nudge that shows how far you've come, and rewards a "100%" state that genuinely benefits the user — an ethical nudge, not a dark pattern. 🎯 Takeaway: the line between a helpful nudge and a manipulative one is "whose goal does it serve?" AI products cross that line easily when they optimize engagement metrics against user interest.

Kodak and the digital camera. Kodak invented the digital camera and clung to film anyway — a failure of intellectual flexibility and the sunk-cost fallacy, not a lack of evidence. 🎯 Takeaway: the incumbent's trap is emotional attachment to a winning bet. In a field moving as fast as AI, the willingness to say "the ground shifted, so we should too" is a survival skill.

Air Canada's chatbot invented a refund policy (2024). A support bot confidently stated a refund policy that didn't exist. A tribunal made the airline honor it. 🎯 Takeaway: the canonical product-sense-for-AI failure — a confidently-wrong output with no grounding, no "I don't know," and full authority in the UX. The fix was architectural: answer only from real policy, with citations. That's a product decision before it's an engineering one.

The "expert's trap" in domain-heavy products. Marty Cagan's warning that experts assume they are the user shows up everywhere from EMRs to developer tools. 🎯 Takeaway: domain expertise is a turbocharger only if paired with the outsider question "why is it done this way at all?" — the same humility that keeps an AI PM validating instead of trusting the model's (or their own) confidence.

Field notes from working PMs

Three durable heuristics from PM-interview collections (Carl Shan's Product Manager Handbook interviews with PMs at Google, Twitter, Facebook, Yammer):

"What is the user's goal — and how hard does this product make it?" (Jason Shah, Yammer). The two-question quality model that needs no dashboard: book a good place fast, get a car fast, find the answer fast. 🎯 Takeaway: before any metric review, walk the product as a user with a goal and count the friction — it's motivation theory compressed into a gut check.

High-quality products do one thing exceptionally well (Avichal Garg, Facebook). Spread across many domains, you're replaceable. Exceptional at one, you're a category. 🎯 Takeaway: the same focused-strategy discipline, stated as a quality bar. It's doubly true for AI products, where the reliable frontier is narrow.

Product sense is built by dissection (Garg again). Extreme attention to detail — dissect what makes something good and what makes something not good — is how pattern-matching intuition is trained. 🎯 Takeaway: pairs with the talk-to-your-users mechanism: exposure supplies the patterns, deliberate dissection compresses them into taste.

Module recap

The through-line: product sense is not one talent but five habits that compound: reading behaviour, simulating the user, choosing a focused strategy, communicating it, and grounding it in domain reality. For the AI PM, a sixth habit sits on top. The model is a new kind of material — powerful, probabilistic, and confidently wrong. The product is the trustworthy system you build around it, not the model itself.

Walk-away question: "For the job in front of me, what does the user actually want to achieve — and if I'm reaching for a model, does it genuinely serve that job better than a simpler, more predictable solution?"

Test yourself

Five questions; answers fold out. If one stumps you, the link takes you back.

1. A feature has high motivation and a strong trigger but users still don't act. What's the first suspect, per the behaviour equation?
   

   Answer

   Ability — friction. B = M × A × T: if two factors are present, the third is the bottleneck. Hunt the needless steps before adding more motivation. (Motivation & behaviour)

2. Why does "would you use this?" produce worse data than "tell me about the last time you…"?
   

   Answer

   People are unreliable futurists but decent historians. Statements about hypothetical behaviour barely predict real behaviour. Accounts of past behaviour do. (User research)

3. What's the one-question test for whether a nudge is ethical?
   

   Answer

   "Whose goal does it serve?" A nudge aligned with the user's own goal is help. One serving only the metric is a dark pattern. (Motivation & behaviour)

4. Your AI feature demos beautifully. What does "taste becomes evals" say you must do before trusting it?
   

   Answer

   Turn your judgment of "good" into a graded, representative eval set and run it continuously. Read traces and call pass/fail yourself. A distribution can't be eyeballed from a demo. (Product sense for AI)

5. When is deep domain expertise a liability, and what's the antidote?
   

   Answer

   When it becomes "I am the user" and validation stops. The antidote is pairing expertise with the outsider question — "why is it done this way at all?" — and continued user contact. (Domain expertise)