blog — 12 entries
On why programming skills fade, what actually brings them back, and how this thing is built.
latest
Reviewing a colleague and reviewing a model are different search problems. One rewards a model of the author, the other rewards a model of what you failed to say — and neither improves by resolving to read more carefully.
5 min read#code-review#ai#practice
5 min read
The scheduler is SM-2, the intervals are 1, 3, 8, 20, 50, 125 days, and the pass/fail that drives it comes from running your code rather than from you rating yourself.
#spaced-repetition#learning#practice
5 min read
Five exercises came out of the generator and all five passed every static check we have. Then we ran them. Four did not work.
#ai#tooling#quality
5 min read
Similarity scoring and keyword matching grade the shape of an answer. A test suite grades the behaviour, accepts any correct solution, and hands back a stack trace — which is the part that teaches.
#testing#infrastructure#quality
4 min read
The bottleneck is not prompting. It is deciding what to build, checking what came back, and knowing when to stop steering and read the code yourself.
#ai#practice#code-review
4 min read
Generating code is now cheap and reviewing it is not. The failure mode is not bad code that looks bad — it is wrong code that looks right, and reads fluently, and passes the tests you thought to write.
#ai#code-review#practice
4 min read
Every exercise had been read and looked fine. Executing the reference solution against the real test suite found eleven that nobody could have solved, including three the author was sure about.
#testing#quality#tooling
4 min read
A container shares the host kernel, so a kernel bug is a host compromise. A microVM does not. Here is what changed when BlankCode moved submission execution from Docker to Firecracker, and what it cost in latency.
#security#sandboxing#infrastructure
4 min read
A test that calls a function and never checks the result passes forever. Reviewing for that by hand does not scale, and it did not work. A forty-millisecond rule found both.
#tooling#testing#code-review
4 min read
A production bug where every server-rendered page died with the wrong exception, two confident diagnoses that were both wrong, and the mistake in the reproduction that let them survive.
#debugging#nuxt#postmortem
4 min read
Skills you built over years fade in months once you stop using them. The fix is not re-reading tutorials — it is being asked to recall the thing right before you would have forgotten it.
#learning#spaced-repetition#practice
4 min read
Turning a procedure you know into instructions an agent follows correctly is a testable engineering problem, not a writing exercise. The test is the case you held back.
#ai#tooling#process