A Flip Is a Missing Rule
A colleague sent me the cleanest bug report I've had all year. He'd taken an open change, squashed it, and opened it again — same content, byte for byte, as far as the reviewer could tell.
Tagged evidence
-
-
Intermittent Competence Looks Like Coverage
My reviewer caught a deleted flush call three times out of three, and said nothing about eight of thirteen removals in any run. A check that works a quarter of the time doesn't present as a gap. It presents as a feature.
-
A Constraint You Can Quietly Lift Isn't One
I had a harness measuring whether a set of skills helped a cheap model. It came back negative: no lift worth the tokens. I believed it.
-
Train on What the Gate Throws Away
My code reviewer does far more work than it ships. Scouts cast a wide net — anything that might be a violation.
-
A Big Rule Set Is an Attention Problem
Sixteen minutes and sixteen dollars to surface four low-severity nits, most of it spent scanning for rules rather than judging code. A corpus can fit the context window and still overflow the model's attention.
-
Rules Are the Alpha
Two reviewers, the same commit — a fat change, a few thousand lines. One was a deep reviewer I own, running a portfolio of judgment rules tested across a whole corpus of real PRs and code.
-
Familiarity Isn't Trust
I caught myself telling an AI, "you're starting to sound like me." My cadence, my hedges, the figures I reach for — it had picked them up and was handing them back. It's a nice feeling.
-
Shrink, Reorganise, Relocate: Why Synthetic Wins Fail on Real PRs
I spent a stretch trying to make an AI code reviewer cheaper without letting it miss more. Cost and recall, two dials pulling against each other.
-
Augmented Reading
The reflex, when a model meets something long, is to make it shorter. Summarise the PR, TL;DR the paper, have the agent review the diff and hand back the findings. It all points one way: read less.
-
You Can't Grade Your Own Misses
Ask a code-review tool how good it is and you'll get a precision answer: of the things it flagged, how many were real. That's easy — every flag is a claim you can check.
-
Fidelity Was a Vibe, Cost Was a Number
I told the agent fidelity came first. It optimised for cost anyway, for an embarrassing reason — cost had an estimator and a guardrail, and fidelity had a sentence.
-
Your Merged-PR History Is Survivorship-Biased Training Data
In the Second World War, the Statistical Research Group was asked where to add armour on bombers, and handed the damage data from the planes that came back — most shot up on the wings, fuselage, and tail.
-
Noticing Is Cheap, Judgment Is Scarce
The goal was never a cheaper reviewer. It was finding everything that is really there while posting nothing that isn't — and that objective, not the budget, is what forced the pipeline apart.
-
Evals Aren't the Moat
A fast eval told me a restructured module was clearly better. Comfortable margin. I ran it again to be sure, got the same verdict, promoted the new version, and moved on.
-
A Pool of Models Narrows the Overlap — It Doesn't Erase It
Sakana AI shipped Fugu with a clean pitch: one endpoint that routes your task across a pool of models — calling itself recursively — and handles selection, delegation, verification, and synthesis internally.
-
The Productivity Mirage
Sixteen experienced developers finished 19% slower with AI and believed they were 20% faster. The feeling of a fast session tracks fluency, not distance.
-
Your Codebase Is the Prompt
The AI reads from your codebase before it writes into it. Whatever's already there — your conventions, your shortcuts, your best work and your worst — is the prompt.
