A second trap was choosing the most attractive number. A CPU percentage from a trace told part of the story, but it did not explain the end-to-end rate. Wall-time counters around the actual transfer path did. Once those existed, the bottleneck moved from a vague system impression to a specific stage we could test.

That changes how I want agents to work. A dashboard is evidence only when we understand what it measures, what it perturbs, and what it can lose. The useful loop is source event to completed outcome, with counters at the boundaries and a second measurement path when the first one is suspect.

The wider lesson is not limited to embedded systems. Logs, traces, browser audits and model evaluations are all instruments. If the instrument changes the work or hides part of it, the right response is not a more confident explanation. It is a better sensor.

The safer shape was simple. Production stayed where it was. A separate review address served the candidate build from the same repository, with an unmistakable review banner. Stakeholders had one stable place to inspect, while the public address remained untouched.

That separation also made the checks useful. Playwright could crawl the sitemap on desktop and mobile, verify canonical URLs, headings, image alternatives, overflow, browser errors and share controls. Lighthouse could inspect the local static build without a review banner contaminating the result. The checks found ordinary defects — metadata gaps, image weight, fonts and hierarchy — before they became production defects.

I now see the review URL as poka-yoke: a small design choice that makes the wrong action harder. Review infrastructure is part of product quality when it protects the live surface and gives humans a calm, repeatable place to decide.

That would have made the archive look tidier, but it would not have been true. The writing first appeared elsewhere. The site adaptation happened now. Both facts matter.

The pattern we settled on is explicit: retain the original publication date and exact source link, show the adaptation date on the page, and use that adaptation date as dateModified in the structured data. Keep the substance and first-person stake, but do not pretend the current site hosted the original.

Good metadata is not only for search engines or preview cards. It is a compact statement of provenance. Honest chronology, clear ownership and approved imagery make an archive more trustworthy than a perfectly smooth history.

The first proper review found that the inbox stopped on launch day. The week itself had been full: hardware evidence, safer communication boundaries, review infrastructure, accessibility checks and a substantial pass over the agent working relationship. The capture mechanism had recorded none of it.

That is useful Hansei. A written rule is not yet standard work. It has to survive the week when attention is elsewhere. The correction is not to generate posts from activity counts or dump a transcript into public. It is to sweep the inbox, weekly git history and public artefacts at EOW, then mark each candidate promoted or skipped.

The empty inbox earned this note because it exposed the gap between intention and operation. Chop wood, carry water: improve the mechanism, publish only what teaches, and leave the rest out.

v1.0.0 is the lock: don’t casually rewrite the public skeleton mid-week. Improve the pack, patterns, and notes when the week earned it. Cut v1.1 when the shape itself changes.

Release and baseline note live on GitHub. Steal the pack. Measure your own results.

We’ll update chopwoodcarrywater.uk in the end-of-week review — the same Hansei pass where patterns get keep / amend / retire. If the week earned a note, a pack tweak, or a pattern change, it ships then. If not, we skip. No filler calendar.

That’s the cadence: weekly with the review, not whenever a chat feels productive.

There’s now a timed checklist and a downloadable Cursor pack: two always-on P0 rules, one skill to rename, one signed pattern, one clock. About an hour. Then stop for the day.

Come back for Toyota Way Practices and Preloop when the boring basics are already earning rent.

On product repos I wire a Preloop Cloud pull-request reviewer. OpenRouter is how that reviewer reaches models. Each push gets a GitHub commit status under context `preloop`. Actions can wait on that status before self-hosted jobs start, or leave the review async when I only want the signal.

It is not merge authority. It is not unit tests. It is shift-left: catch the semantic miss before the runner eats the afternoon. Keys stay vaulted. Humans still merge.

The temptation after a good insight is novelty — another skill, another abstraction, another brand. The useful move is usually the ordinary one: keep the sensors green, keep the gates honest, keep the always-on budget thin.

So when we needed a proper domain for this notebook, I didn’t want agent-something-ai.com. I wanted the proverb. The .com was taken; .uk was free. That felt right.

We keep a lean core (memory, decisions, app control, a small durable runtime layer) and enable specialty tools only for the job. When the job ends, unload. No “leave it on in case”. Reloading the window is cheaper than carrying last week’s AppSec server forever.

If a session feels heavy, that’s a bug. Measure it. Cut it. The working relationship stays useful when the rent stays boring.

What stuck for us: when something actually works — or fails twice the same way — propose a thin pattern, get a human sign-off, encode it where agents will load it. Kaizen with teeth. Hansei when it hurts.

That’s why this notebook talks about skills, rules, and signed patterns instead of “prompts”. The model is powerful. The operating system around it is what makes Monday morning repeatable.