The factory

In April I started a monorepo called Clawset with one question: how far can you get if agents do the building and the human only makes decisions? It's what I do in the evenings. The day job pays for the habit.

Version one: a factory with a board

Ten loops ran as scheduled Claude Code routines in the cloud, each with its own prompt, cadence and job:

• a brainstormer that wrote new ideas against a set of shared themes
• an idea fleshouter that researched one idea at a time and scored it on seven dimensions, with a sentence of reasoning for each score
• triage, which promoted the good ones and kept the build queue full
• a scoper that turned an idea into a plan of 12 to 30 small tasks
• a builder that ran four times a day and took one plan to a working site in a single session, one pull request per task
• an executor for follow-up work, an auditor that read one project a day and filed what it found, and a validator that fixed whatever broke
• nightly stats that wrote me a digest and watched every loop's heartbeat, and a weekly planner that wrote a retro and regroomed the board

Every merge landed in one daily batch that type-checked, tested and deployed only what changed. A dashboard rebuilt itself after each batch, so I could review the whole thing like a board: steer the queue, freeze a project, unblock a loop.

Between May and August the loops wrote 1,012 run logs, merged 4,782 pull requests and deployed 144 sites. One busy day in July saw 176 commits. On a good week the planner's retro counted around twenty new live sites.

What it taught me

Volume stopped being the hard part almost immediately. Three lessons stuck.

Measure from day one, or don't pretend to. The first design killed sites that missed their traffic targets by a deadline. Analytics were never wired into the loops, so the killer was judging on no data at all. In July I turned it into a build machine instead and made kill decisions mine.

A knowledge base compounds only where you feed it. Ideas hung off eight shared themes, and the brainstormer was meant to work the thinnest one. Five themes never got real research notes, so it kept mining the same three. The loop was right to refuse; the gap was mine.

Picking what to build is the job. Most of the output was programmatic SEO. The idea that held up was a buyer's guide for secondhand gear.

Version two: one job at a time, all the way down

That idea became allbutnew, with its own repo and a different way of working. The scheduled loops have been paused since August.

Two skills do the work. /ticket writes the job up and sizes it. /ship takes it from commit to a verified live deploy, with a review that scales with the change and a second model reviewing alongside. A small tower view tells me which of my terminals is waiting on me. The rules fit on a page: decide and keep going, stop only for a cloud bill going up or data that can't come back.

The write-up has the details and the prompts to build your own.