How this backlog got built

01 Source

The backlog started as a single markdown table: a slug, a topic, a source, an interaction idea, a status, one row per piece. The rule from the top was no invented filler — every row traces to something that already existed. A shipped pull request. A saved note. An actual git log. Where the source material was thin, the instruction was to write around the gap rather than manufacture specifics to fill it.

02 Draft and build

Each row became a job with the same shape: pull the real source, draft it in voice, fit it to a shared kit of layout primitives — a prose-and-specimen unit, a labeled flow diagram, a before/after toggle, a sidenote — then build it. Several rows ran at once, each handed to an agent in its own isolated git worktree, so one agent's half-finished edit couldn't collide with another's mid-build. The one file every worktree agreed to leave alone was the registry: the single list of which pieces exist, wired in by hand after each batch landed, specifically so parallel agents never had to resolve a conflict in the same file.

03 Verify

Verification was mechanical and non-negotiable before anything shipped: typecheck, lint, tests, a production build. One more check sat on top of those, specific to this site: grep every draft for the brand names of AI products, because a personal site tied to a day job at a company with its own model can't casually name a competitor's. It caught something real. One of the twenty drafts named a specific LLM by brand, three separate times — the source material it was pulled from had quoted the name verbatim, and the draft carried it straight through. The grep caught it before merge, in the same pull request. A separate, site-wide sweep run around the same time found a couple of stray mentions already live elsewhere and pulled those too.

04 Integrate

The registry file was supposed to be the one place parallel work couldn't collide. It mostly held. Then two independent batches of agents, started from two different sessions neither aware of the other, built the same four pieces at the same time — same slugs, same arguments, one batch naming a file with a lowercase final letter and the other with an uppercase one, two on-disk names for one essay. Nobody caught it mid-build, because nothing in either worktree could see the other. It surfaced at integration, when a second batch's git-worktree listing came back already full of work that looked exactly like its own. The fix wasn't a smarter agent. It was a habit: check what's already running before starting a new batch, not after.

The table near the top of this piece is real, and it's already wrong by the time you read it — a few of the rows marked queued there will have shipped between the draft and this sentence rendering on your screen. That's not a bug in the tracker. A backlog built by a swarm of parallel agents doesn't have a single moment when it's fully described anywhere; it has a table someone updates by hand, a registry file that's the actual source of truth, and a gap between them that closes and reopens with every batch.