you are not the sea

essays on AI, leverage, and organizations in the age of agents. pick a lane in AI, Code, Productivity, or People. or read my Highlights. find out about me and subscribe.

Nov The Expensive Minute. 200 PRs a day was supposed to buy me leisure. Instead each minute got more expensive, and the idle ones cost the most. Keynes promised the fifteen-hour week; Linder explained why it never arrives. , So Why Not Spend $2,000? If tokens track output, 100 engineers at $1,000 a month of tokens is roughly 50 at $2,000. Of course you spend the $2,000. What makes it hard is that it's people and livelihoods. And if it holds, why not keep the 100 and double the output instead? , The Cold Read A PR description is a claim, and whoever reads it first is primed by it. So one agent describes every PR from the diff alone, a second grades how well that agrees with the original description or prompt, and the grade tells automerge (and me) how far the description can be relied on. , My AI Coding Workflow (Revisited) A fresh snapshot of how I'm working with AI coding tools, nine months after the first one. The flow is mostly the same; what changed is how much of it I no longer touch. , Getting Wider Working inside other organizations as a forward deployed engineer, the AI adoption curve I see is getting wider: at the front each step makes the next one cheaper, at the back each step not taken makes the next one more expensive. Cloud and git closed that gap eventually; I don't know whether this one does. , Tunnel Vision A support document for a half-built feature came back with "not implemented yet" scattered through it. Claude and Codex both prop drill, write tests for 1+1 and over-disclose, and most of it is tunnel vision: the task in hand feels like the most important thing in the world. My partial fix is separate steps in separate contexts.
Oct The Importance of Backpressure Kanban cards were backpressure in physical form: a fixed number in circulation, so upstream could not outrun downstream. Software kept the board and dropped the pressure. , One Hundred, and Fifty I hit 100 merged PRs in a day this afternoon, not long after turning 50. Both are round numbers, and both are more useful as questions than as trophies. , Build Your Own Lightsaber Vibe code your own harness tools. Like building your own lightsaber, it teaches you how the thing actually works, and it should be cheap enough to throw away. , More Attention, Not Less I set a goal of 100 PRs a day because I was drowning in context switches. The target forced me to treat my own attention as the bottleneck. , Don't Run It Twice Most of what CI does on an agent's PR is repeat what the agent ran ten minutes earlier. So the agent leaves a git trailer listing the checks it already ran, and the PR build skips those. That was a good third of taking ten-minute builds to ninety seconds, and main re-runs everything anyway. , Start With One Tell someone an agent merged to production with no human in the loop and the reply is "that's not possible." Nobody is proposing the general case, so ask whether one is possible instead: mine was a minor patch bump. , Four to Five Days Claude sized a feature at four to five days, unprompted, then we shipped it in a few hours. It estimates in human days because everything it knows about estimating came from humans with calendars. A quirk, mostly, and a reminder that I don't size work in days anymore.
Mar The first PRs shipped by Customer Success. Last week our customer success team shipped their first two pull requests. It changed how we ship and how we communicate. , Battleship Prompts Same task, three different prompts. See which one gets closest. It's surprisingly effective and makes you a better prompter. , A Demographic Collapse of Developers Junior developer hiring is stalling. Meanwhile, the senior pipeline doesn't refill itself. , The Claude Dichotomy AI is creating a genuine dichotomy: teams drowning in AI-supercharged chaos, others are thriving. The difference is in the approach to transformation. , Tool Augmented Prompting Your linter, compiler, and test runner already talk to the AI. Make them say something useful. , Do as I say not as I TODO Your AI assistant holds a mirror up to your codebase. 5 ways to make sure it reflects the good parts. , No agents, no plan, ship it Why I prefer single-prompt one-shotting with AI coding tools over multi-step agents, swarms, and planning modes. , We're not vibe engineering Vibe coding is the headline. Engineering a new SDLC is the actual work. , Clear Your AI Mind Your context window is not a diary, don't forget to /clear. , Spin the Bottle Neck The bottleneck in software has moved. Teams are still optimising for the old (wrong) one. , Still choose boring technology Dan McKinley's classic advice on "choosing boring technology" is even more true in the AI era. , AI Prompts are Wishes AI prompts are wishes. The more powerful the wish, the bigger the potential for the curse. , Skills I'm ceding to AI My shortlist of coding skills I'm genuinely comfortable handing to AI. , Don't Believe the Hype AI vibe coding tools aren't truly there yet. But early adopters can see they will be. Those aren't the same thing.
…

See all articles, highlights, or find out about me and subscribe.