Decisions, trade-offs, and what held up in production. Filter by category, or read everything.
AI made it cheap for a small team to offer more, and most are spending that on ground someone already covers. Why widening looks like Google's graveyard, and why cheaper execution is a reason to go deeper.
Maintainable code was insurance against the cost of changing it later. Models collapsed that cost, so the backend can carry plain, improvable code, and the cleverness belongs where the customer meets the product.
Looking at work and seeing how it could improve feels like the same act as seeing what it takes to ship. It is the opposite act. Count the steps, box the effort, and cross the line.
Menus were a workaround for software that could not understand a sentence. That constraint is gone, and most SaaS interfaces have not noticed. What changes when the product knows the user's next N tasks and walks them through.
Running an autonomous agent on someone else's data is the sharpest form of embedded engineering. The failure modes that surfaced, and why each fix moved the discovery out of the customer's run.
A partner was told a dashboard clone was queued. Minutes later it failed on something we could have checked before answering. How the pre-flight replaced the apology.
When your product runs on the customer's data, in their workspace, sometimes in their cloud, done stops meaning what your test suite says. Four rules, each learned by breaking it.
Partial success is the dangerous path. How a clone gate, a precise cache flush, and one side-by-side proof kept a customer's first open from being a dead screen.
A tile choice turned a contractor's three-centimetre error into a bill. The rule underneath: tolerance is a budget somebody pays, and precision only earns its cost where errors compound.
Reversible-first thinking says reach for the cheap option before the heavy one. It is incomplete: a cheap attempt is only cheap if failing it leaves the expensive attempt exactly as available, and exactly as cheap.
Both disciplines describe a system precisely enough for something else to execute. What separates them is the price of being wrong — which sets how many attempts a career contains, and explains most of what people mistake for culture.
A dream is a plan with feelings attached, and plans get audited. Price its ceiling, tear it down if it cannot pay, rebuild on the outcome you actually wanted — the clock is the one input you cannot get more of.
The 3D work was good and the ceiling was real. Tearing it down was not giving up on the dream — it was keeping the dream and demolishing the vehicle that could not carry it.
When agents write the code, your leverage moves to defining what done means — and building the verification system that checks it. A product builder's guide to the shift.
No fireplace, one energy source, isolated utility rooms, redundant solar, protected structure. What twenty years of software and DevOps did to a house design.
A small, un-opinionated agent skill that impersonates any persona and walks any user story against a live product. What it catches, what it can't, and why it's read-only.
Social housing failed partly because uniformity was the price of scale. Generative design makes personalization nearly free on paper — which reopens an old question.
Doing a task by hand improves the task. Doing it through a skill improves every future run — self-documenting, delegable, verifiable, and harder to skip steps in.
Mass production serves the median and abandons the rest. The use of 3D printing in building is the low-volume, high-variety part nobody will tool up for, not the whole structure.
Most advice treats AI as an accelerator for what you do well. The bigger win is documentation, runbooks and process discipline: the work you always skip.
Teams were promised a 3-10x multiplier and report closer to 10%. The gap isn't the model; it's that most codebases give an agent nothing to check against.
Agreeableness is a dial, not a personality. Set it too high and you watch avoidable damage happen in silence. Too low and people stop bringing you things.
Impact-effort ranking assumes your tasks are independent. They aren't. Order work by what it makes cheaper next, and by what still matters after a pivot.
Handing over output is fine. Handing over the decision about what's worth making is how people become interchangeable. A practice for keeping the difference.
The cost of a house is a stack of numbers, each improvable by a few percent, and they compound once design iteration gets cheap.
Generative design stalled in architecture because it needed a programmer on staff. That just changed, and form optimisation is still barely explored.
Some problems can be planned. Others only show their shape once you're inside them. How to tell which you're in, and what AI changed about finding out.
The chances that change a trajectory are rare, short and badly timed. You cannot schedule one. You can be able to move when it opens, and size the move so a miss does not end you.
My grandfather sold what he grew. My father sold what he built, locally. I work for people I'll never meet. The bottleneck moved from land to judgment.
Remote work removes every ambient signal people used to reassure themselves with. The fix isn't monitoring. It's written process and unambiguous ownership.
Joining after two calls, and what stuck: why the process existed before I arrived, two opposite ways to find a bug, and how to point your paranoia at customers.
I had a defensible position and I was pleased with it. Then someone ran the numbers. On the gap between having a considered view and feeling like you have one.
Nothing in this category yet. Read everything instead.