The Occasional Jam No. 09

No. 09 soapbox

Where the Code Comes From

3 min read

Most of the code on this site is written by an AI.

I build JamCatWow in the margins of my week, a handful of hours, most of them late. That only adds up because I stopped writing most of the code. I decide what gets built, and I review what the AI writes.

The coworker with amnesia

My collaborator starts every session with no memory of the last one. Whatever we decided yesterday is gone unless it's written down where the next session will look. So the repository does the remembering.

Flowchart: five boxes feed into a fresh session with a blank memory. The boxes are the house rules (always true), the glossary (what words mean), decision records (what we chose), stream records (where it stands), and the code and its tests (what is true). Flowchart: five boxes feed into a fresh session with a blank memory. The boxes are the house rules (always true), the glossary (what words mean), decision records (what we chose), stream records (where it stands), and the code and its tests (what is true).
What the robot reads before it types.

A glossary pins the product's words, One game's puzzle for one specific date., Anyone playing a game. The glossary also lists the banned synonyms: user, customer, member, and visitor., and The one code path that uploads a finished puzzle. Everything published crosses it, so nothing ships unchecked. among them, because a collaborator that uses three names for one thing will build it three ways. Decision records, one short file per settled argument, keep old debates closed.

The house rules live in one file the AI reads before it touches anything: never an em dash in player-facing copy, the production database never gets ahead of the code, and the version number goes up in the commit that earns it.

It's the onboarding doc you'd write for a new teammate, except the teammate is new every morning.

Guardrails that answer in seven seconds

The AI makes mistakes faster than I can review them, so the checks have to be faster than me. Before every push, a hook runs the whole test suite on my Mac, and the verdict lands about seven seconds later. The robot reads the same verdict and fixes what it broke before I look.

Some of the rules started as accidents. A newer version of the package manager rewrote the file that pins exact dependency versions and broke the deploy, twice, so now there is a rule about which install command the robot may run.

When something burns us, the fix becomes a sentence in the file the robot reads first.

The rules for what a day's puzzle looks like are written twice, on purpose: once in the Python that publishes a day, and once in the TypeScript the site reads it with. Two copies of a rule drift. By July 2026 they disagreed in three places, and nothing had noticed. A day the publisher accepts and the site refuses means players get a practice puzzle instead of that day's, and nothing logs it.

None of the three had broken a real day yet. Now 132 example days run through both copies on every push, and if the two give different verdicts on any of them, the push is blocked.

The publisher's copy of the rules

Example dayPublisherSite
A normal dayacceptsaccepts
A word with an accent in itrejectsrejects
A date written 12345-1-01rejectsrejects
An answer that is blankrejectsrejects

The copies agree on all 132 examples. The push goes through.

By July 2026 each of these three rules was looser in one copy than the other, and nothing was checking.

One robot per worktree

For a while every session worked in the same checkout of the project, and one afternoon two of them collided. A branch switch in one swept the other's half-written work into an unrelated commit and reverted files out from under it. Nothing was lost. That was the end of sessions sharing a folder.

Now every stream of work gets its own A separate copy of the project folder that shares one git history, so two sessions never edit the same files., and the main checkout stays clean for merging. The four setup steps git won't do for you have a page of their own.

The robot sounds like a robot

The prose was harder than the code. First drafts of the tour post were accurate and read like a press release with jokes: a punchline at the end of every paragraph, lists that come in threes, questions that answer themselves. We named the tells together and wrote them into the style rules, and now the robot checks its own drafts against them.

It drafted this post too, against those rules, and I edited it.

The parts that stay mine

The AI does most of the typing, most of the first drafts, and the audits I'd never have the hours for. It doesn't want anything.

It has no opinion about whether a game is fun, and it will build a bad idea beautifully.

So I decide what gets built, and I decide what stays. A new game gets four weeks after launch. If nobody I know is playing it five days a week by then, it comes off the site.