Loops
Agents on a schedule, with a human approval gate
Skills are capabilities: instructions Claude Code follows when I invoke them. Loops are what happens when a skill runs on a schedule. An agent does real work against real data and brings me a proposal. Nothing merges or deploys itself; every loop ends in the same place, me reviewing a branch. This page documents the loops currently running on this site.
growth-loopEvery Monday morning an agent reads this site's analytics, filters out the bot noise, and looks for one copy-shaped problem: a page people leave too fast, a headline that promises the wrong thing. It forms a falsifiable hypothesis, rewrites the words on a branch, verifies the build, and writes me a plain-English report: the problem, the hypothesis, the change. I approve or reject. The next run scores the result.
- Copy only: no layout, styles, or components
- One small change per run
- Never pushes, merges, or deploys itself
- “Nothing worth changing this week” is a valid outcome
- A human approves every merge
site-updatesTwice a month an agent reads every commit since the last time the Project journal page was curated, clusters them into themes, and writes the story (what was built, why it was needed, and what it changed) in plain English, extending an existing arc when the work continues one. The new entry lands on a branch with the build verified, and I read the story itself as the approval request. Raw commit logs never reach the page.
- Stories, not commit digests: the validator rejects hash dumps
- At most two entries per run
- Never pushes, merges, or deploys itself
- “Nothing worth recording yet” is a valid outcome
- A human approves every merge