https://eshpel.in/blog Product, writing, design, and engineering essays from Saquif Alam. en Saquif Alam Tue, 25 Aug 2026 08:20:25 GMT https://eshpel.in/blog/flyover-always-feels-faster https://eshpel.in/blog/flyover-always-feels-faster Sun, 16 Aug 2026 00:00:00 GMT AI We started the last year with reading articles that METR had measured experienced developers going 19% slower with AI while they believed they were going 20% faster. The more interesting thing happened in February 2026, when METR admitted it could no longer run the study, because it could not find enough developers willing to spend a day working without an AI tool. Saquif Alam https://eshpel.in/blog/agent-user-cannot-file-a-bug https://eshpel.in/blog/agent-user-cannot-file-a-bug Sun, 16 Aug 2026 00:00:00 GMT PRODUCT 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 :/ Saquif Alam https://eshpel.in/blog/praise-to-criticism-ratio https://eshpel.in/blog/praise-to-criticism-ratio Thu, 09 Jul 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/the-agent-doesnt-get-a-seat https://eshpel.in/blog/the-agent-doesnt-get-a-seat Tue, 16 Jun 2026 00:00:00 GMT AI For fifteen years software was sold by the seat, one named user for one monthly fee, back when every seat was a person and the software only made that person faster. Then the agent started doing the work itself, and a single agent that clears what seven hundred people used to clear is an agent that empties seven hundred desks you were billing for. Sell that by the seat and your revenue shrinks in step with how well the thing works. Saquif Alam https://eshpel.in/blog/ship-an-mcp-or-become-a-database https://eshpel.in/blog/ship-an-mcp-or-become-a-database Sun, 07 Jun 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/desire-paths-are-the-product https://eshpel.in/blog/desire-paths-are-the-product Tue, 26 May 2026 00:00:00 GMT 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. Saquif Alam https://eshpel.in/blog/build-to-learn-vs-build-to-earn https://eshpel.in/blog/build-to-learn-vs-build-to-earn Tue, 19 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/wanting-the-wanting https://eshpel.in/blog/wanting-the-wanting Mon, 18 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/margin-notes-on-spiegel https://eshpel.in/blog/margin-notes-on-spiegel Mon, 18 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/why-i-stopped-reading-product-substacks https://eshpel.in/blog/why-i-stopped-reading-product-substacks Sun, 17 May 2026 00:00:00 GMT REFLECTION Eight subscriptions ago, my product thinking was a compressed mirror of whatever Marty Cagan and Lenny Rachitsky had written that month. Then I unsubscribed, one at a time, over a single quarter. The thinking that came back was sharper, and the trade is one I am quietly recommending to anyone who will sit with it. Saquif Alam https://eshpel.in/blog/why-everyone-in-dhaka-has-a-side-hustle https://eshpel.in/blog/why-everyone-in-dhaka-has-a-side-hustle Sun, 17 May 2026 00:00:00 GMT REFLECTION Two colleagues running two economies on two monitors between sprint meetings. One is QAing a Saturday build for an agency client, the other is approving a Daraz refund. The day job is the visa stamp. The side hustle is the actual job, and that has been true in Dhaka since roughly 2018. Saquif Alam https://eshpel.in/blog/english-medium-school-ipo https://eshpel.in/blog/english-medium-school-ipo Sun, 17 May 2026 00:00:00 GMT REFLECTION Annual tuition at a name-brand Dhanmondi English-medium school sits well above what a mid-career private-sector parent can comfortably afford, and yet every gate has a waiting list. The argument here is that the school is no longer selling education; it is selling a long-dated option on emigration, priced to extract every taka the parent will not admit they are spending. School names below are deliberately fictionalised. Saquif Alam https://eshpel.in/blog/dont-trust-the-user https://eshpel.in/blog/dont-trust-the-user Sun, 17 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/bkash-agent-is-the-bank-branch https://eshpel.in/blog/bkash-agent-is-the-bank-branch Sun, 17 May 2026 00:00:00 GMT REFLECTION bKash didn't win on app polish or frontier ML. It won by realising the bottleneck on Bangladeshi financial inclusion was a trust handshake, and the handshake had to happen at a corner store with a laminated pink poster and a cash drawer. The next ten consumer categories in this country need the same retail-shaped on-ramp. Saquif Alam https://eshpel.in/blog/bangladesh-is-the-worlds-best-ab-testing-ground https://eshpel.in/blog/bangladesh-is-the-worlds-best-ab-testing-ground Sun, 17 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/support-team-is-the-eval-set https://eshpel.in/blog/support-team-is-the-eval-set Thu, 14 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/notes-on-cat-wu https://eshpel.in/blog/notes-on-cat-wu Thu, 14 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/margin-notes-on-karpathy https://eshpel.in/blog/margin-notes-on-karpathy Thu, 14 May 2026 00:00:00 GMT REFLECTION I watched Andrej Karpathy's vibe-coding talk with my own blog open in another tab. Most of what struck me wasn't argument, it was overlap. Five places where his sentence in San Francisco met a paragraph I'd already pencilled in Dhanmondi, plus the one place I'd push past where he stopped. Saquif Alam https://eshpel.in/blog/have-we-confused-thinking-with-programming https://eshpel.in/blog/have-we-confused-thinking-with-programming Thu, 14 May 2026 00:00:00 GMT CRAFT I passed a senior engineering interview last month and I still couldn't write a React calculator from scratch. The grammar of code is not the thinking of code, and a generation of Dhaka programmers, like a generation of Dhaka grammar-exam toppers, has been confusing the two for thirty years. Saquif Alam https://eshpel.in/blog/casino-is-the-land-registry https://eshpel.in/blog/casino-is-the-land-registry Thu, 14 May 2026 00:00:00 GMT REFLECTION Two colleagues wear the same tight pleasure on their faces, one with a Crash bet slip on his phone, the other with a Purbachal deed on his desk. The argument here is that the only honest difference between the two is the timeline, and that Bangladesh has been pretending otherwise for thirty years. Saquif Alam https://eshpel.in/blog/streaming-is-ux https://eshpel.in/blog/streaming-is-ux Tue, 12 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/slack-is-real-org-chart https://eshpel.in/blog/slack-is-real-org-chart Tue, 12 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/pricing-ai-by-token-trap https://eshpel.in/blog/pricing-ai-by-token-trap Tue, 12 May 2026 00:00:00 GMT AI A Karwan Bazar fish caller has watched six pricing units come and go in thirty years, and the SaaS bhais selling AI agents are about to discover the per-token unit moves under them the same way the maund did when the ice started melting. The only shape that holds is per-outcome, and nobody has worked out how to bill it yet. Saquif Alam https://eshpel.in/blog/migration-silent-pm-job https://eshpel.in/blog/migration-silent-pm-job Tue, 12 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/half-life-of-benchmark https://eshpel.in/blog/half-life-of-benchmark Tue, 12 May 2026 00:00:00 GMT AI MMLU was a real measurement in 2023 and a polite formality by 2025. HumanEval, GSM8K, MATH all went the same way. A benchmark is alive only as long as the gap between models still maps to something a customer can feel; once the labs train against it, the rate card on the wall becomes decoration. Saquif Alam https://eshpel.in/blog/coder-taught-model-to-ship https://eshpel.in/blog/coder-taught-model-to-ship Tue, 12 May 2026 00:00:00 GMT CRAFT Vibe coding is real but mis-described. The actual skill is the older one of reading a diff and saying which lines stay. Compiling is the floor a tabla apprentice clears in week one of riyaz. The ceiling is taste, learned the way the apprentice learns it, by being corrected on the same bol until the wrist stops cheating. Saquif Alam https://eshpel.in/blog/brief-as-prompt https://eshpel.in/blog/brief-as-prompt Tue, 12 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/type-schemas-not-tokens https://eshpel.in/blog/type-schemas-not-tokens Mon, 11 May 2026 00:00:00 GMT AI Most agentic codebases treat the model's output as a string to parse. The clerk at counter no. 3 of Gulshan post office knows better. Schemas are the form, and tokens that don't match the form go nowhere. Saquif Alam https://eshpel.in/blog/read-twenty-roadmaps-write-one https://eshpel.in/blog/read-twenty-roadmaps-write-one Mon, 11 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/rag-retrieves-wrong-thing https://eshpel.in/blog/rag-retrieves-wrong-thing Mon, 11 May 2026 00:00:00 GMT AI Six specific failure modes I keep meeting in production RAG over brand-voice corpora, in roughly the order I find them. Most teams discover them after launch, because their eval set was built from happy-path queries that never asked the awkward question. Saquif Alam https://eshpel.in/blog/pm-stack-prompt-eval-observability https://eshpel.in/blog/pm-stack-prompt-eval-observability Mon, 11 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/multilingual-is-a-moat https://eshpel.in/blog/multilingual-is-a-moat Mon, 11 May 2026 00:00:00 GMT CAREER Old Dhaka, off Tipu Sultan Road, a signboard painter who has been mixing enamel for forty years explains the part of PM work nobody puts on the JD. Translating between PRD, SQL, customer-voice, engineering-speak and the boardroom is the same job as choosing whether the next letter should be brush-stroke Bangla or stencil English. Saquif Alam https://eshpel.in/blog/hiring-for-taste https://eshpel.in/blog/hiring-for-taste Mon, 11 May 2026 00:00:00 GMT CAREER Tools have flattened output. Anyone can produce a wireframe, a sales email, a roadmap. The moat that's left is taste, the curated set of what's worth shipping and what's worth a polite no. Here's how to test for it in 45 minutes. Saquif Alam https://eshpel.in/blog/evals-are-release-notes https://eshpel.in/blog/evals-are-release-notes Mon, 11 May 2026 00:00:00 GMT AI The PR description has stopped being the changelog. The pass-rate movement on the eval suite is. v2.3 lifted retrieval-grounded quotes from 81% to 89%; that one line is the release note. Everything else is the annual-day speech. Saquif Alam https://eshpel.in/blog/death-of-prd https://eshpel.in/blog/death-of-prd Mon, 11 May 2026 00:00:00 GMT PRODUCT 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. Saquif Alam https://eshpel.in/blog/compounding-context-cost https://eshpel.in/blog/compounding-context-cost Mon, 11 May 2026 00:00:00 GMT AI Every agentic call drags the whole running history back into the prompt, and the meter on that ride is quadratic. Teams stop noticing because each individual hop feels cheap, the way a CNG passenger stops watching the meter once Mohakhali clears. Saquif Alam https://eshpel.in/blog/career-advice-half-life https://eshpel.in/blog/career-advice-half-life Mon, 11 May 2026 00:00:00 GMT REFLECTION Most career advice has a shelf life of roughly eighteen months. A retired Sylheti headmaster spreads forty years of notebooks on the clothesline after the monsoon, reads what dried, what curled, what buckled. Six lines that didn't make it, and one rule that did. Saquif Alam https://eshpel.in/blog/caching-is-hospitality https://eshpel.in/blog/caching-is-hospitality Mon, 11 May 2026 00:00:00 GMT ENGINEERING The cache is a boarding-house register, not a hash map. The bua of a four-room mess in Sutrapur explains who she lets stay, who she evicts before noon, and why LRU is a landlady's instinct that the Netflix caching team eventually wrote down on a slide. Saquif Alam https://eshpel.in/blog/bun-is-the-new-node https://eshpel.in/blog/bun-is-the-new-node Mon, 11 May 2026 00:00:00 GMT ENGINEERING A team switching JavaScript runtimes is usually less dramatic than the standup that decided to do it. The interesting part of moving from Node to Bun isn't the runtime, it's that the audit forces you to confront which of your dependencies were quietly lying about being standard the whole time. Saquif Alam https://eshpel.in/blog/build-is-the-brief https://eshpel.in/blog/build-is-the-brief Mon, 11 May 2026 00:00:00 GMT ENGINEERING Compile errors and broken tests are the most honest user research you will ever get from a feature. The PRD only tells you what the team wishes were true, the build only ships what is actually true, and the fastest way to close the gap is to let the build read the spec back to you. Saquif Alam https://eshpel.in/blog/ai-eats-qa-first https://eshpel.in/blog/ai-eats-qa-first Mon, 11 May 2026 00:00:00 GMT AI Everyone has been worrying about whether models will replace engineers. The role they have been quietly merging with for the last eighteen months sits one floor down, counting colonies on agar plates at 7am, and the technician is still on the seven o'clock bus. Saquif Alam https://eshpel.in/blog/agentic-contract-tests https://eshpel.in/blog/agentic-contract-tests Mon, 11 May 2026 00:00:00 GMT ENGINEERING A unit test on a deterministic function asks for an exact byte. An agent answers in a different shape every Tuesday, and the byte changes. What you actually want is a contract on what the answer must look like, witnessed by three named clauses, the way a kabin-nama gets signed in a Wari sitting room. Saquif Alam https://eshpel.in/blog/good-things-are-good-things https://eshpel.in/blog/good-things-are-good-things Sun, 10 May 2026 00:00:00 GMT REFLECTION A friend's old-money-versus-new-money list landed on my feed and got two hundred likes by lunch. What the list is actually doing, more quietly than the sociology it's pretending to be, is laundering virtues through a class label nobody questions until you point at it. Saquif Alam https://eshpel.in/blog/tests-are-the-spec https://eshpel.in/blog/tests-are-the-spec Thu, 07 May 2026 00:00:00 GMT AI Specs rot in ninety days. Code, in an agentic codebase, rewrites at the speed of pull requests. The only artifact that can't rot is the test suite, because CI runs it on every push and a stale test fails. That changes whose job it is to write them. Saquif Alam https://eshpel.in/blog/agentic-ab-test https://eshpel.in/blog/agentic-ab-test Thu, 02 Apr 2026 00:00:00 GMT AI Classical A/B testing assumes humans branch. When both sides of the marketplace are agents, what's the control group? Saquif Alam https://eshpel.in/blog/privacy-is-the-product https://eshpel.in/blog/privacy-is-the-product Thu, 12 Mar 2026 00:00:00 GMT 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. Saquif Alam https://eshpel.in/blog/rag-field-guide https://eshpel.in/blog/rag-field-guide Sun, 15 Feb 2026 00:00:00 GMT AI Lessons from building retrieval against brand-voice corpora at Fortune 500 scale. What actually breaks, what to evaluate, and why retrieval is the easy part. Saquif Alam https://eshpel.in/blog/diagram-vs-document-pm https://eshpel.in/blog/diagram-vs-document-pm Wed, 04 Feb 2026 00:00:00 GMT CRAFT Two PMs face the same problem. One reaches for a whiteboard. The other reaches for Notion. The artifact you start with sets the ceiling on what you can ship. Saquif Alam https://eshpel.in/blog/on-successfully-failing https://eshpel.in/blog/on-successfully-failing Thu, 22 Jan 2026 00:00:00 GMT REFLECTION What seven years and one wound-down EdTech taught me about the kind of quitting that compounds. Why dissolving with grace beats dragging the corpse around. Saquif Alam https://eshpel.in/blog/roadmap-compression https://eshpel.in/blog/roadmap-compression Thu, 08 Jan 2026 00:00:00 GMT PRODUCT Leadership sees four boxes. The team sees forty. What ships is three, plus an unrelated thing. We're not behind. We're compressed. Saquif Alam