Topic · PRODUCT
Writing on product management craft
Field notes from running platform and integration product teams at enterprise scale, with a hands-on engineering bias.
21 posts in this topic. See all writing.
The phrase “the roadmap is a compression algorithm” came out of a Tuesday review in 2025 where seven quarterly themes had to fit on one slide for a Board meeting in eleven hours. That essay is the spine of this cluster. Around it sit “The death of the PRD-as-document,” “The PM who reads twenty roadmaps and writes one,” “Migration is the silent product job,” and “Wanting the wanting,” which is a margin note on Stephen Spiegel that turned into an essay on feature-request hedonic loops. The argument the cluster makes, taken together, is that the artifact a PM ships is almost never the artifact the PM is graded on. You ship the PRD but you’re graded on the migration. You ship the roadmap and you’re graded on what got compressed out of it, and by the time you’ve written the spec the engineers are grading you on the empty middle between weeks two and seven instead. Read in sequence, the eighteen pieces give you a working theory of what a Principal Product Owner inside an enterprise marketing-software company actually does between Monday standup and Friday demo. The AI co-pilot in the IDE has already eaten half the old job description, and the other half is still being negotiated quarter by quarter.
- Aug 2026PRODUCT3 min
Your agent user cannot file a bug.
What happens when our product has acquired a user that reads your error strings, acts on your defaults, and never once complains. It does not answer surveys, it does not open tickets, and when it gets confused it just retries :/
- Jul 2026PRODUCT7 min
Count both queues.
A tailor can't unit-test a punjabi, so he judges his Eid season by who walks back into the lane afterwards: the customers who come back carrying new cloth, and the ones who come back with the shoulders folded in a bag. AI products have put the rest of us in the same position. When you can't fully verify the thing before it ships, the praise-to-criticism ratio in your feedback stream becomes the closest thing you have to a report card, and it's a better one than NPS ever was.
- Jun 2026PRODUCT7 min
Ship an MCP or become a database.
Greenhouse shipped an MCP server this year, and the part worth sitting with is not the feature list. It is that a recruiter can now run a full hiring report, pull candidate context, and clear a batch of rejections without ever loading greenhouse.com. The interface layer of SaaS is migrating into the agent, and the product you spent a decade making beautiful is becoming a backend somebody else narrates.
- May 2026PRODUCT8 min
Desire paths are the product.
A planning officer in Lisbon paused mid-walking-tour last spring in front of a fresh limestone plaza and pointed at the dirt line cutting diagonally across the middle of it. The same failure mode happens in product, where the team confuses the polished surface it shipped with the path the user actually wore. The cleanest products of the last thirty years are the ones that figured out which one was the artefact and which one was the data.
- May 2026PRODUCT9 min
Build to learn vs build to earn.
Marty Cagan's distinction between build-to-learn and build-to-earn has been getting passed around in product Slacks for a year now, mostly read as advice for early-stage teams. The argument I want to add, after a quarter of running the experiment myself on an established B2B marketing-workflow platform with thousands of paying enterprise marketing teams attached to it, is that the biggest scope for build-to-learn in 2026 is sitting inside a working build-to-earn product, not in the lean-startup garage everyone keeps drawing it in.
- May 2026PRODUCT8 min
Wanting the wanting.
There is a stand-up bit I caught on a podcast where the comic points out that you don't want new things, you want the wanting, because the moment a thing arrives in your hand you start hunting for the next thing to almost-have. The B2B PM version is that customers request features they then never use. The request itself was the value, and the feature sits in production while the same customer files the next one.
- May 2026PRODUCT9 min
Margin notes on Spiegel.
Evan Spiegel sat down with Lenny and said, roughly, that software stopped being a moat fifteen years ago and that the people running scared from AI now are running scared from that same fact in different clothes. I watched the interview with a half-finished bKash post still open on the next tab, and most of what struck me was how cleanly the Dhaka examples already half-argue his point.
- May 2026PRODUCT9 min
Don't trust the user.
A B2B customer wrote in to say our one-time codes were five characters long when the field on his screen only took four, and I spent half a Tuesday in the production database hunting for a bug that turned out to be a column he had not scrolled past. The wider claim is that watching a user does more work than talking to one ever will, and that in B2B SaaS a complaint without a session replay attached is half a complaint at best.
- May 2026PRODUCT9 min
Bangladesh is the world's best A/B testing ground (and nobody's using it).
Bangladesh has 170 million people pressed into a strip a tenth the size of the US, segmented across language, religion, income, age, and which phone the user is holding, riding a unified payment rail. It is the cleanest signal-to-noise ratio in any consumer market on earth, and almost no global product team runs experiments here.
- May 2026PRODUCT8 min
Your support team is the eval set you're not using.
In agentic-AI products, the people answering tickets and walking new customers through their first ninety days already hold the failure-mode catalogue your eval team is trying to write from scratch. The under-priced research function is sitting one floor down, badged in a different colour, and the company keeps treating it like a cost line.
- May 2026PRODUCT9 min
Cat Wu would not hire most of you.
Cat Wu runs product for Claude Code at Anthropic. She told Lenny Rachitsky, plainly, that most PMs she interviews are approaching the role wrong, that her team prefers engineers with taste, that PRDs are mostly gone, and that some features ship in a day. I watched the interview with my own /blog open in another tab and threaded comments down the right margin.
- May 2026PRODUCT7 min
Streaming is UX, not plumbing.
From a Banani 11 cha stall, a small argument about why token streaming is the most underrated UX primitive of the decade. The act of watching the characters appear changes how the user reads, edits, and trusts the response, and a product team that mistakes the pour for plumbing is selling cha by handing the customer a thermos.
- May 2026PRODUCT6 min
Your Slack channels are your real org chart.
The printed reporting structure is a fiction. The real org is who DMs whom, which channels stay alive, which ones go quiet. A traffic constable on the Mohakhali flyover reads the road the way a senior engineer reads a Slack workspace, and the painted lines are not the lanes.
- May 2026PRODUCT9 min
Migration is the silent product job.
Most product writing is about new features. The unwritten skill is moving fifty thousand users from v1 to v2 without losing any of them. The actual job lives in the labels on the boxes, the way Mintu bhai does it on a Friday afternoon in Banani.
- May 2026PRODUCT7 min
The product brief as a prompt.
When the brief is what generates the implementation, the brief becomes the spec, the test, and the documentation collapsed into one artifact. A film-set script supervisor in Old Dhaka has been doing this for forty years with a 40-page screenplay in one hand and a one-page call sheet in the other. The crew shoots the call sheet.
- May 2026PRODUCT7 min
The PM who reads twenty roadmaps and writes one.
Twenty inputs land on the desk by Friday afternoon. Customer interviews, sales notes, support tickets, three rival roadmaps, a legal risk register. The job isn't planning. It's de-duplication, prioritisation, and the kind of grammar that lets a tired team find the spine of next quarter.
- May 2026PRODUCT7 min
The new PM stack. Prompt, eval, observability.
Roadmaps and PRDs are still around, but the day-to-day for AI-product PMs runs on three rails. The prompt is the artifact, the eval is the test, and the trace is the truth-teller. If you can't read a Langfuse trace, you can't ship.
- May 2026PRODUCT8 min
The death of the PRD-as-document.
The Product Requirement Document was a 1990s artifact for waterfall org charts. In 2026 the artifact is the eval set, the prompt, and the trace. PRDs still get written, but past page two nobody is reading. The new spec is executable, and a stamp doesn't make a document; it certifies one.
- Mar 2026PRODUCT6 min
Privacy is the product.
Most products treat privacy as a constraint. A few make it the entire value prop. The inversion changes everything about how you build, market, and price.
- Jan 2026PRODUCT6 min
The roadmap is a compression algorithm.
Leadership sees four boxes. The team sees forty. What ships is three, plus an unrelated thing. We're not behind. We're compressed.
- Oct 2025PRODUCT5 min
Half-life of a spec.
Day 0 the spec is god. Day 30 we mostly followed it. Day 90 the spec is folklore. Day 180 the team is making fun of you for caring.