I spent a week marking a model's homework: did it catch the decision I actually made? Slow, tedious work. Then, closing it out at 7:25 on a Friday night, a bug: every trace had been cut to only its first line. The model had seen a fifth of what I'd said. You can't be serious? There was a lot of swearing.

A preread for our catch-ups, a mirror of a shared board, a screen, a sidecar, a watcher. Each one worked, once. Looking again, it's horribly over-engineered. Less is less. They're not worth their maintenance. I'm retiring more code than I'm creating now. It feels good.

Fifteen days and 3.6 billion tokens on one curious question. Nothing else, 'til it was answered, understood, solved, and shipped. Then, everything else, all at once. Fish tacos, a burst pipe, a new switch in the wall, and a jig in millimetres.

Not how most people's brains work, I don't think.

278 tasks done. 126 open. 23 killed.

The 23 is the interesting number.

Trello holds the tasks and Google Docs holds the writing, with a small script in the middle moving things between them. Auth, comments and notifications are somebody else's solved problem, done better than I'd manage and thrown in free. The join is the part that's actually mine, so that's where the time goes.

I wanted dry air for spraying lacquer. That became a second-hand industrial air dryer — a thing I hadn't known existed when I started — sixty brass fittings from two specialty suppliers, and a weekend project to install it.

A music server leaks memory, taking about a week before freezing the host machine. I could chase this upstream, through other people's code, for an unknown number of evenings. Winning is choosing not to. A nightly 4am restart gets the job done.

Text is either bold or it isn't, which turns out to be one option short. I wanted the weight in between — semibold, 600 — and the bold button only makes 700. So now the renderer writes the weight itself and I never touch the button. Nobody else is ever going to notice.

Left seven comments on a document generated by my own agent, then sorted them by which layer had failed instead of just fixing seven things. The document was the test case. The comments were the report.

An agent that reads email and decides what's worth recording agreed with its human reviewer 27 times out of 27 over a fortnight.

It is still not allowed to decide anything on its own.

Spent an afternoon in Fusion 360, drawing a 3D model for a welding table, because nothing I could buy was right. Now I know arbitrary things about manufacturing in steel with lasers. In the end, I bought something off the shelf.

Building got cheap, which makes it very easy to solve problems nobody actually has. So I don't start until there's real work waiting for the thing. That settles the biggest question first — whether anyone wants it — and running it on something real always turns up a second version that's better than the one I would have imagined.

Most leaders are focussing on tools and platforms. The ones pulling ahead are closing loops.

read the article

Pipelines put a human at the finish line taste-testing whatever the machine produced. That's a workflow, not a loop.

read the article

Every feature ships easily, and none of them compound. It feels fine at the start. By feature 10, the product feels incoherent. Every new addition makes integration harder and more fragile.

read the article

AI didn't change where design starts. It changed which layers are cheap. The work didn't disappear — it moved, and the layer where it landed is the one most teams treat as somebody else's problem.

read the article

We can now ship working software faster than it takes to decide whether it is the right thing to build. The technology did not break — the collaboration model did.

read the article

They have already thrown resources at AI and hit a ceiling. Not a technical ceiling — a capability one, which is much harder to name and much harder to buy.

read the article

My brain was a riot of distraction. You become a jack-of-all-trades, master of none, in an environment that rewards depth and expertise, not breadth and adequacy.

read the article

Ask ten product managers what their job actually is, and you'll get eleven opinions. Trying to define Product Management is the wrong approach — you'll just get academic answers that don't help anyone do the work.

read the article

It's not about speed anymore. It's about wisdom when everyone can build anything, anytime.

read the article

We can 10x what we build. We cannot 10x our customers' attention. They make up their mind once, and they do not forgive a wasted visit.

read the article

There's nowhere to hide in a document. It's the facts and narrative that tell the story, not the energy of the person speaking, or the quality of their over-designed slide deck.

read the article

Even successful workshops typically produce transient information — it's only meaningful if you were there — and the Miro board or walls of sticky notes make no sense a few days later.

read the article

No product playbook, no design training. They had lived the customer's pain many times, in different ways. They were experts on the problem. That was the part that mattered.

read the article

They built their own support platform when every sensible piece of advice said buy one. That is conviction: building what customers need before they know they need it.

read the article

When there's even a hint of ambiguity about the company's goals and priorities, teams insert their own interpretations that favour their pet projects or hobby horse topics.

read the article

Strategy is hypothetical while leadership is human. A strategy document makes assumptions about the future. A leader interprets real-time signals from the market and adjusts accordingly.

read the article

The operating model that got a founder this far is the one stopping them going further. It worked, which is exactly why it's so hard to change.

Testing a hundred ideas does not buy a hundred ideas' worth of certainty. It buys ninety-nine cheap ways to be wrong instead of one expensive one.

The number nobody quotes back is the ninety-nine.

Design Thinking is how we explore and solve problems; Lean is our framework for testing our beliefs and learning our way to the right outcomes; and Agile is how we adapt to changing conditions with software.

read the article

Interviewing is a quest for learning, not a quest for certainty, so the consequences of being 'wrong' are relatively low. You don't need a certification to be an empathetic listener.

read the article

Insight that arrives weeks or months after the question is often irrelevant. Not because it is wrong, but because the team has already moved on to the next question.

read the article

Software engineering is not merely execution of a plan, it also defines strategy.

read the article