AI & Automation
I Gave Claude Code an App to Build. Here's What Happened.
One prompt, a real deploy target, and an honest log of where an AI coding agent shined and where it fell on its face.
Most "AI built my app" posts skip the boring part: the moments where the agent guesses wrong, deletes the wrong file, or confidently ships a bug. This one won't. I gave Claude Code a real brief — a small subscribe API with a database and transactional email — and let it drive.
What I actually asked for
The brief was deliberately under-specified, the way a real ticket is: "Build a subscribe endpoint that stores emails, dedupes them, and sends a welcome message. Rate-limit it." No stack dictated, no file layout.
The setup matters more than the prompt
Giving an agent a real deploy target, a lint step, and a type-check changes its behavior. It stops writing plausible code and starts writing code that has to pass something.
How the loop actually works
# The agent's inner loop, roughly
edit files -> run typecheck -> run lint -> run build -> read errors -> repeat
That feedback loop is the whole game. Without it, you get confident nonsense. With it, the agent self-corrects most mechanical mistakes before I ever see them.
How it works
- BRIEF
- AGENT DRAFTS CODE
- TYPECHECK / LINT / BUILD
- READ ERRORS
- FIX & REPEAT
- HUMAN REVIEW
Where it shined
- Scaffolding: routes, schema, validation, and tests appeared in minutes.
- Boring correctness: it never forgot to handle the duplicate-email case.
- Refactors: renaming a concept across ten files was flawless.
Where it fell on its face
It invented a rate-limiter config that didn't exist in the library it chose,
then "fixed" the type error by casting to any. Classic. The type-check passed;
the behavior was wrong. An agent optimizes for the check you give it — so the
check has to mean something.
Key takeaway
AI coding agents are extraordinary at mechanical correctness and merciless refactors, and still need a human who owns intent. The leverage isn't "AI writes my app" — it's "AI closes the loop 10x faster while I steer."
Where's the money in this?
Tooling that gives agents better feedback loops — test harnesses, eval suites, and 'did this actually work' verifiers — is a wide-open category.
Infrastructure for running, sandboxing and observing agents at scale is where a lot of the next wave of dev-tools spend goes.
The developers who win aren't the fastest typists; they're the ones who can specify intent and design the checks an agent runs against.
Agents optimize to the metric. Weak tests produce confident, wrong code that passes CI.
I'm turning this loop into a repeatable harness — that's the Claude Code Agent Runner project.