What do I use every day?

The tools change often, so what matters is the job each one does.

  • Claude Code, for changes across several files, refactors and writing tests.
  • Cursor, for editing in one file with the rest of the project in view.
  • The Claude API inside the products themselves, for the AI features a client actually asked for.
  • Plain editor and plain terminal, for the parts where I need to see exactly what is happening.

How does a build actually go?

The plan comes first and is written by me, because the model has never met you and does not know what your business can tolerate.

  • I write the plan and the acceptance conditions: what the screen does, what happens when something fails, what must never happen.
  • I build the shape of it myself: the data model, the routes, the boundaries.
  • AI fills in the repetitive middle, and writes the tests for it.
  • I read every line before it goes in. If I cannot explain a line, it does not ship.
  • I run it on a real phone or a slow connection, which is where most problems appear.
  • You see it working early, on your own device, and the plan changes in writing if it needs to.

Where does AI code go wrong?

The failures are predictable, which is what makes them catchable.

  • Invented details: a function or a setting that sounds right and does not exist.
  • Confident handling of the wrong edge case, usually the one that matters most to you.
  • Happy-path code that assumes every request succeeds and every field is filled in.
  • Security that looks fine until you ask what happens with someone else's record.
  • A tidy refactor that quietly removes behaviour nobody wrote down.
  • Slow queries that only appear once there is real data in the table.

The rules I keep

These are not slogans, they are the checks that stop the failures above.

  • I read and understand everything that ships, and I am the one accountable for it.
  • Tests are written for the parts that break quietly, and run after every change.
  • Credentials and customer data never go into a model prompt. Your data reaches an AI feature through the product's own API call, with the account in your name.
  • Anything a customer reads gets a human approval until the error rate is acceptable.
  • Every AI step keeps a log of what it was given and what it returned, so a wrong answer can be traced.

What it changes for you as a client

More gets done inside the same fixed price, and more of my time goes to the parts that still need a person: your process, your edge cases and testing.

It also changes what is worth building. Things that used to cost too much for a small business, like an internal tool or an automation with an AI step in it, are now within reach. Small jobs still start at $50.

What to remember

  • AI writes code, a person owns the decision.
  • Read everything, test the parts that fail quietly.
  • Keep credentials and customer data out of prompts.
  • Log every AI step so a wrong answer can be traced.

Sources

Primary documentation, for anyone who wants to check the details.