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.

cyclegrowth-loop
ActiveWeekly — Monday mornings

Every 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.

Each run:
1Pull analytics2Filter bot noise3One hypothesis4Rewrite on a branch5Verify the build6Report7Approval8Measure next week
Guardrails:
  • 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
Built on:ga-reportgrowth-loopScheduled task; also runs on demand