Notes
Notes
Technical and operational point of view. Durability over hype.
We scored our own site, then fixed it: a changelog
The same rubric we used on 771 agencies, pointed at ourselves. What it found, what we changed, and how the score moved from 11th to 1st on web design.
Read the note →Fix the loop, not the output: why our build pipeline heals itself
When an AI build fails, the tempting move is to patch that one result by hand. It will fail again on the next build. A case where three fixes to the loop took a stuck build from 84% to 97%.
Read the note →The state of agency websites in October 2026: speed, accessibility and AI readiness
We measured 726 agency sites with Lighthouse, axe and a crawler. Median mobile performance is 56, nearly nine in ten have a serious accessibility finding, and almost nobody publishes in two languages.
Read the note →How we grade an agency website: the 36-criterion rubric
Six criteria are measured by script and thirty are judged, each from 0 to 20. The scale, the rules that keep judging consistent, and what we left out on purpose.
Read the note →Why a software studio should publish its prices
In our benchmark of 771 agency websites, price transparency had a median score of 5 out of 20. The sites that publish a range or a model are the ones that stand out. What we publish and why.
Read the note →Silent failures: the bug shape that costs us the most
Loud failures get fixed in an hour. The expensive bugs are the ones that report success. Four patterns we have hit while running our own products, and the one rule that catches them.
Read the note →Building a benchmark harness that survives its own failures
To score 808 websites we ran headless browsers, Lighthouse, accessibility scans and AI judges for hours. The first run died at site 327 without an error. What we changed so a failure costs minutes, not the whole run.
Read the note →We used two AI models to judge 771 websites. What they got wrong.
Absolute scores flattered us, head-to-head ranking was more honest, and the two judges only half agreed. What we would do differently if you copy this method.
Read the note →What the top 10% of agency websites do differently
We compared the best-scoring tenth of 771 agency websites with the rest. The gap is specificity, not speed or polish: concrete AI stories, published prices, clear process and security answers.
Read the note →We benchmarked 771 agency websites, including our own. Here is what we found.
We measured and judged the websites of the Clutch 1000 against 36 criteria. The field is slow, short on proof, and quiet about price. Our own site ranked 11th on design and 571st on proof.
Read the note →Turning a PDF dump into a public atlas: how we built Atlas Criminalidad CR
A country’s crime data was open and unreadable at the same time. Here is the architecture that turned a quarterly stack of PDFs into a tool anyone can read in 5 minutes — and that keeps working every release.
Read the note →The five-minute rule: why slow WhatsApp response is your most expensive problem
Every minute a lead waits for a reply is a minute they spend deciding whether to try someone else.
Read the note →Optimize for year three: why most software dies in year two
Building for the third year is a set of unglamorous decisions made on day one.
Read the note →The teardown list
One engineering teardown at a time — real systems, real failure modes, no marketing. Unsubscribe any time.