1. Measure before claiming

    A performance or quality claim needs a before number, an after number and the method. I report per screen or per case rather than one flattering average, and I say what the measurement does not cover.

    React Native work reported each screen's React Profiler timing, labelled as development-build numbers. See the case study

  2. Let checks decide, not the model

    When an agent writes code, its own "done" is not evidence. I put the decision in something that cannot be talked round: a test suite, a lint rule, a visual-diff threshold.

    Ralph Harness keeps a commit only when your checks pass; the design-system toolkit blocks off-system imports before the file is written. See the case study

  3. Options, cost, recommendation, then build

    Before a technical choice on a client project I write a short comparison: the options, what each costs, what I recommend and why. The client decides with the trade-offs in front of them.

    Client choices of providers and services start as a written comparison with prices and a recommendation.

  4. Report what did not work

    Status updates and write-ups include the failures and the limits. It is cheaper for everyone to hear early that a metric misleads than to find out in production.

    The Figma-to-code article has a section on where screenshot comparisons misled us. See the case study

  5. Own it end to end

    I am comfortable owning the whole path: the code, the deploy, the server, the support inbox and the pricing page.

    Site Spy: I build it, host it, price it and answer its users. See the case study

  6. Safety net before the risky change

    Large migrations go first through tests that pin down current behaviour. The migration then becomes a check against known results instead of a leap.

    On the trading platform, the hooks and exchange logic got tests before the MUI v5 to v7 upgrade. See the case study