We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1) Big Ideas
AI should widen the option set, not make the product decisions. Ravi Mehta defines taste as curation: knowing the possibility space, deciding what matters, and choosing what is right for an intended audience. His AI workflow is frame → explore → critique → explore again → select → refine, replacing ask → generate → accept; use AI aggressively for exploration and cautiously for decisions. For PMs, generate several interaction or feature directions, then record the audience, constraint, and evidence behind the selection. When your preference and user preference diverge, interrogate the gap rather than dismissing the user.
The harness is part of the product. Hiten Shah argues that even a great model fails inside a product that “loses the thread.” Harness-Bench ran 5,194 agent trajectories across 106 tasks and saw average harness scores from 52.4 to 76.2 with the same task suite and model pool. Treat context continuity and agent evaluation as product requirements, not post-launch plumbing; improve the surrounding system before assuming the model is the bottleneck.
2) Tactical Playbook
Make metrics actionable, not ceremonial. A Reddit synthesis of two 2026 product surveys reports that 34% of PMs have no clear primary metric, 49% lack time for analysis and tracking, and 40% do little or no experimentation; it also reports that only 4% think company metrics reflect their work. Treat this as directional: the thread flags sample bias and KRs that PM teams cannot independently move, particularly in B2B. Start with one leading indicator believed to connect to revenue or retention, track it consistently, and replace it if evidence breaks the link. Leadership-set roadmaps can be appropriate when executives have better customer context; the test is customer evidence, not whether leadership was involved.
Qualify enterprise demand before allowing it onto the roadmap. A bank-and-insurer vendor’s checklist is concrete: find the P&L owner with budget and mandate, request the security questionnaire by the second meeting, and attach a budget line to pilots. Multi-thread above and beside a champion, and make one-off custom work paid. This filters interested-but-powerless buyers and prevents unpaid, customer-specific scope from becoming product strategy.
Try the smallest fix before rebuilding. A team with unusable Metabase image thumbnails considered a new internal tool, then used a Chrome extension to zoom on hover in five minutes; the dashboard stayed fast and the team shipped faster. Isolate the missing capability, test a reversible workaround, and only then price the operational cost of a rebuild.
3) Case Studies & Lessons
Cursor made a contrarian product-layer bet, then changed layers without losing the thesis. It focused on the human–model interface rather than a coding-specific foundation model or VS Code plugin, because users express intent in natural language and a plugin would make Cursor part of someone else’s product. It treated product quality as go-to-market and deferred enterprise. As models advanced, it moved from IDE to agent platform to model platform in about two years; user scale supplied data and know-how for its own models, and later enterprise economics justified a rapid sales build. Lesson: define the layer you can own, let it constrain scope, and revisit the layer when technology or margin structure changes.
4) Career Corner
AI PM hiring has expanded while general PM hiring contracted. Aakash Gupta’s tracking shows 16,420 AI PM listings versus 35,725 PM listings—46% of open PM roles, up from 2% in February 2024—while total PM jobs fell 18%. In a hand-classified sample of 113 postings, 69% asked for prior AI/ML experience, but only 14 required production or scale experience; median PM tenure was five years while examples often asked for one or two AI years. The author notes that 40 classifications were judgment calls. The practical opening is to build a real AI product, get it used, document decisions and evaluation artifacts, and publish it as verifiable work—not fabricated experience. One example: after 15 years in Apple hardware and a senior program role, Brian Luc built RenoSmarter.ai and became an AI PM at Cisco five months later.
5) Tools & Resources
Use a structured feedback request. Product communities’ Feedback Friday template asks for the company, URL, purpose, technologies, specific feedback requested, beta-testing need, and additional comments; it discourages link-only promotion and asks feedback seekers to return feedback. Reuse it for user research or beta recruitment: a concrete question and reciprocal ask should produce more actionable input than “thoughts?”
