Rhys Brett-Bowen Coding in the age of AI

Start here

There are 43 posts. These are routes through them, in order. Each one follows a single argument rather than collecting everything on a subject — the archive and the tag filter already do that.

Start here

The argument in three posts: what turned out to be scarce, what it cost to build for that, and why I can't prove it worked.

  1. Give Away the Writing, Keep the Reading one public repository, read closely; the guard and the caching catch are experience and a wager
  2. Noticing Is Cheap, Judgment Is Scarce metered runs, mine; two claims marked unproven in the text
  3. You Can't Grade Your Own Misses real planted-bug runs that establish a floor, not superiority; the corpus that would settle it does not exist

Building an AI code reviewer

The receipts, in the order the problems arrived. Each post is a thing that broke and what I changed.

  1. Noticing Is Cheap, Judgment Is Scarce metered runs, mine; two claims marked unproven in the text
  2. A Big Rule Set Is an Attention Problem lived, plus a small directional check (7B, n=6 per cell); the recall effect is unproven
  3. Fidelity Was a Vibe, Cost Was a Number dollar figures metered and real; the projections are my own estimator, not a billing oracle
  4. Rules Are the Alpha the rule portfolio is real and tested; the transfer check is one 7B model at n=6, directional only
  5. Train on What the Gate Throws Away the pipeline is real and running; training a model on the pile is the part I have not done
  6. The Harness Is the Product the harness is real and running; repointing it to a new domain is argued, not shipped
  7. Keep the Factory, Ship the Result the pipeline is real and running; the factory-versus-output framing is my argument from building it
  8. A Pipeline Is What It Refuses to Forward the pipeline is real and running; that it improved anything beyond cost is an untested hypothesis

Whether to believe the machine

How to tell a working check from one that works a quarter of the time, and why your own evaluation is the last thing you should trust.

  1. You Can't Grade Your Own Misses real planted-bug runs that establish a floor, not superiority; the corpus that would settle it does not exist
  2. Intermittent Competence Looks Like Coverage measured — 13 fixtures, 39 runs, mine; corpus unadjudicated, so a shape not a rate
  3. Evals Aren't the Moat cited prior art, including a result that contradicted my instinct and stayed in because of it
  4. Your Merged-PR History Is Survivorship-Biased Training Data a real census of roughly 248,000 closed PRs; everything downstream of it is reasoned prediction
  5. A Pool of Models Narrows the Overlap — It Doesn't Erase It cited research plus cross-family experience; the boundary itself is a wager
  6. Confidently Wrong cited research plus experience; that calibration now beats confidence is a wager

What it does to your own skill

The part that isn't about tools. What you stop practising, what that costs later, and which difficulty is worth keeping.

  1. The Productivity Mirage prior art cited; the gate and the three-asset framing are my synthesis, and a wager
  2. Give Away the Writing, Keep the Reading one public repository, read closely; the guard and the caching catch are experience and a wager
  3. Thinking at the Speed of Code cited expertise research; applying it to code review is a reasoned bet, flagged in the text
  4. Where the Reps Come From low to medium, and split — the retention half is lived, the acquisition half is reasoned
  5. Fluent in a Dead Idiom two real review moments; the mechanism is argument from them, not data
  6. Replace the Trivial, Augment the Difficult cited research plus lived use of a reviewer I built; the synthesis on top is mine

Putting it in front of a team

Where the checks go, what a review has to teach, and the intent no description carries.

  1. How a Team's AI Coding Flow Should Be Wired my model, argued from practice rather than measured; the tooling gap is my inference from our own stack
  2. Teaching the Reviewer cited prior art plus my own experience; that it lifts the team's floor is a wager
  3. The Nit Was Never Optional Google's published standard, cited; that AI reprices iteration is lived, not measured
  4. A Description Can't Replace the Code cited prior art; reading correlated errors onto unstated intent is my wager, flagged in the text
  5. The Other Intent my framing of a split the literature already draws; the synthesis is mine, not the parts

Or read the archive newest first, or start with The Brief, which was an earlier experiment and reads as a story rather than a set of receipts.