Writing on building with AI tools, running a community, and what actually ships.
A stack trace tells you where your program died, not where it went wrong — and the guard clause your agent adds to make the 500 disappear usually deletes the only signal you had. Here's how to trace a bad value back to its source instead.
Agents debug by reading files until they have a theory, and the theory is usually wrong. Here's the alternative: one structured probe at the boundary, a failing and a passing payload, and the actual answer in thirty seconds.
Three unrelated bugs produce the same production 404: missing SPA fallback, case-sensitive filesystems, and static exports that only serve what the build enumerated. Here's how to tell them apart in ninety seconds and a script that catches the worst one before deploy.
Flaky tests feel random, but the research and the debugging both land in the same two buckets: something checked before the work finished, or two things interleaved in an order you never specified. Here's the triage, worked through a real queue-backed invite bug.
Your repo describes maybe half of what your project needs to run. The rest lives in your home directory, your shell profile, and a Docker volume you forgot about — here's how to pull it back into the repo and test it cold.
Your agent writes textbook REST endpoints and one of them will double-charge a customer. Here's how to reason about HTTP constraints — search bodies, idempotency keys, pagination — and write the tradeoffs down where your agent will read them.
The blast radius of an AI coding agent is a property of its credentials, not its intelligence. Here's how to scope database access with real Postgres roles, branches and hooks — and where each layer actually fails.
When your agent's code is "almost right", patching the line you noticed just starts a whack-a-mole loop. Here's the thirty-second diagnostic that tells a bug from a wrong domain model, and the definition-audit-edit repair that actually ends it.
AI coding tools default to a centred spinner because that's the median of every React tutorial ever written. Here's how to fix the four ways that breaks a real table, and how to make your agent stop generating it.
If a test run blows its budget you'll stop trusting it — not consciously, you'll just start typing --no-verify. Here's how we cut a 3m50s Postgres suite to a 1.4s inner loop without buying speed with flakiness.
Free tiers fail in three distinct ways — pausing, gating, and metering — and each one breaks your app differently. Here's how to make each failure visible before a user finds it for you.
A worked example where the "obvious" serialisation fix bought 0% and a one-line index bought 50x. Plus the three ways measurement quietly lies to you: ranking by mean, trusting averages, and profiling somewhere your users aren't.