Things we figured out the hard way .

If you can't see it run, it's not in production.

Three things every system we hand over does on day one: it emits structured events, it redacts what must never be logged, and it signs every state transition. Here's why all three are cheap now and brutal later.

Aug 24, 2026READ →

POSITION3 MIN

The check that stopped checking, and kept saying yes.

From the build log, July to August 2026. Our security self-test went dark for eight days and nobody noticed, because it kept exiting with an error code, which looked like it was working. The lesson is about a whole family of failures: the guard that says done without doing anything.

Aug 17, 2026READ →

OPINION3 MIN

A decision form beats a chat thread.

From the build log, July to August 2026. When a fleet of agents runs overnight, the scarce resource in the morning is not compute. It is the founder's attention. We stopped asking questions one at a time in chat and started batching them into a single form, with a recommendation on every line. Here is why that changed the shape of the whole operation.

Aug 10, 2026READ →

TEARDOWN3 MIN

A lane that goes quiet gets ninety minutes, then it gets stopped.

From the build log. A background process that writes our shared record of the day stopped writing at twenty to eleven. It did not crash. It did not complain. It sat there, alive and doing nothing, while two hours of work queued up behind it. What we changed is not the process. It is how long we let silence run.

Aug 3, 2026READ →

PROCESS3 MIN

The third red changes the design, not the coder.

From the build log, July 2026. A security guard for our code failed review five times, and each time the fix was correct. The failure kept coming back through a different door. The rule we wrote afterwards: when the same class of failure returns three times, stop fixing and start redesigning.

Jul 27, 2026READ →

METHOD3 MIN

The review gate is a machine, not a meeting.

From the build log, July 2026. We stopped asking one reviewer whether code was safe to ship and built a gate that four independent reviewers have to clear before the door will physically open. The interesting part is what happened when we tried to be polite to it.

Jul 20, 2026READ →

CASE NOTE6 MIN

We automated the intake and kept the care human.

From the build log, July 2026. A church we build for collects prayer requests, decisions and connect cards through its website. We put a small machine behind the forms to gather every submission into one clean list, without duplicates, within minutes. Then we deliberately stopped, because deciding what a prayer means and who should follow up is a person's job, not a scheduler's.

Jul 6, 2026READ →

CASE NOTE7 MIN

The service moved an hour earlier, and the page stayed dark.

From the build log, June 2026. A church we build for streams its Sunday service online, and the website is meant to switch itself to live the moment the service starts. One Sunday the service moved an hour earlier and the page stayed dark, not because anything broke, but because the time was written in two places and only one of them got the memo.

Jun 29, 2026READ →

POSITION5 MIN

A book on the shelf has never improved a codebase.

From the build log, June. We read one of the best books on software design and watched it change nothing. Then we did the unglamorous step almost everyone skips: turned it into a one-page checklist, bound to our own code, short enough that a tired reviewer, a junior, or an AI coding tool actually uses it. The reading is the easy part. The translation is the work.

Jun 22, 2026READ →

TEARDOWN8 MIN

The automatic update that took every site we run offline at once.

From the build log, June. One quiet morning every public site we run went dark at the same second, while every health check still read green. The cause was an update nobody chose to run, and a single door that two programs both wanted.

Jun 15, 2026READ →

The machine was excellent. It was pointed at the wrong target.

From the build log. We checked our own AI setup expecting to find a missing tool. The tools were fine. The hold-up was that almost every decision had to pass through one person, and we only spotted it once we counted.

Jun 10, 2026READ →

An agent can do the work and still lose your trust.

From the build log. The agent was doing the work and getting the answer right. The person using it just couldn't see any of that happen, so they stopped believing it.

Jun 3, 2026READ →

TEARDOWN4 MIN

Six small bugs that made a new user's first session feel broken.

From the build log. None of these bugs showed up in a demo. We only found them by signing up as a total stranger and pretending we'd never seen the product before.

May 27, 2026READ →

METHOD7 MIN

The live gate you cannot skip for an agent that pushes code.

From the build log. Before we let a software robot make changes to our code on its own, we rehearsed it against a throwaway copy first. The rehearsal caught two problems that all our careful checks had missed.

May 20, 2026READ →

One command should set up an AI assistant for everyone, or you have not really built a tool.

From the build log. We wanted every person on a team to get their own AI assistant without us writing fresh code for each one. The day the setup tool first ran for a real person, it taught us almost everything worth knowing.

May 13, 2026READ →

The audit that overturned its own brief.

From the build log. We set out to rebuild something we were sure was half-finished. An hour spent actually looking proved us wrong. It saved about two weeks of work on a problem that did not exist.

May 6, 2026READ →

How to give a beginner a capable AI agent without it being reckless.

From the build log. Before handing capable AI agents to a group of complete beginners, we had to settle one question: how much power do you give someone who can't yet spot the danger? The honest answer was not "as little as possible".

Apr 29, 2026READ →

PROCESS4 MIN

Why we ship the agent before we send the invitation.

From the build log. Six welcome emails sat ready to send. We didn't send them: not because of a deadline, but because the thing each email promised wasn't ready yet for the person it was going to. Knowing not to press send was the actual work.

Apr 22, 2026READ →

A fast validator and a slow gate, in that order.

From the build log. We built a content pipeline that writes its own social posts. The interesting part is not the writing. It is the two checks we put between the machine and a human, and why one had to be cheap and one had to be slow.

Apr 15, 2026READ →

TEARDOWN6 MIN

The intermittent error that was a ghost, not a bug.

From the build log. A booking system for a medical-conference client was failing about four in every ten requests from the public internet, while the same request from a laptop worked every time. We spent half a day fixing the wrong thing.

Apr 8, 2026READ →

The production row our code was never written for.

From the build log. We built a feature that gives every trainee their own AI assistant to chat with. It worked perfectly against our test data, then failed against the one real account that mattered most.

Apr 1, 2026READ →

TEARDOWN5 MIN

Restarting one container took down four sites that had nothing to do with it.

From the build log. We restarted one small piece of plumbing to make a routine change. The change itself was harmless. But four unrelated public sites went dark, because of a shared layer almost nobody thinks to check.

Mar 25, 2026READ →

METHOD5 MIN

One team builds client work and trains new people. Keep the two apart.

From the build log. The same small team does private client work and teaches newcomers on the very same platform. Those two worlds must never see each other. The thing that keeps them apart is duller, and far more important, than any password.

Mar 18, 2026READ →

TECHNICAL6 MIN

Stamp identity on the message, not on the context.

When several assistants share one inbox, the question "who sent this?" has to be answered by the message itself. Let the surroundings fill in the blank and you have built something that acts on the wrong person's authority.

Mar 11, 2026READ →

TEARDOWN5 MIN

A merged pull request is not a live deploy.

From the build log. A one-line fix passed every check and merged cleanly, yet the live website kept serving the broken version for five minutes while a real person sat there unable to save her work. Every signal said done. None of them was watching the thing that mattered.

Mar 4, 2026READ →

Building a conference sign-up form on tools we already had.

From the build log. A medical conference needed a sign-up page. The interesting question was not which services to pay for, but whether we could build the whole thing without renting anything new, and what that saves the next client.

Feb 25, 2026READ →

Why event email belongs on an API, not an SMTP relay.

From the build log. The hardest part of a registration system was the bit everyone assumes is solved: sending the "you're booked" email. It looked finished long before it actually worked.

Feb 18, 2026READ →

When a client job is actually a product you needed anyway.

A single medical conference needed a way to take registrations. Halfway through building it, we stopped and asked a harder question: were we building this once, or building the thing our own organisation had needed all along?

Feb 11, 2026READ →

CASE NOTE4 MIN

A whole firm's client database, on a stack that costs nothing to run.

From the build log. A small South African accounting firm wanted one place to see its clients, its fees and its overdue returns. We built it in a single session, put it online, and the monthly bill came to nothing. Here is why that is the right answer for a firm this size, and where it stops being the right answer.

Feb 4, 2026READ →

Why the user accounts existed but nobody could log in.

From the build log. We set up login accounts for a small team the quick way, by writing them straight into the database. Every account showed up in the records. Not one of them could actually sign in.

Jan 28, 2026READ →