← Jamie Hancock

/do

Multi-agent Claude Code workflow · Nov 2026

  • Nov 2026
  • A Claude Code skill
  • Built with Claude Code
This site's plan tree on 5 Nov 2026, near the end of the build. The main screen took 3 rounds to pass; the 5 demos took 1 or 2. ✓ passed, ● in progress, – replaced by a newer piece. r1 is round 1.

/do is the workflow I use for jobs with many parts. I give Claude one line, and it plans the work as a tree of small pieces. I approve the tree once.

Then subagents build each piece in a loop. After every round, a critic agent checks it against one question (one checks copy voice, another demo safety) and sends back what failed.

A piece gets 3 rounds at most. Once a group is done, a critic checks the pieces fit together, and at the very end a fresh critic with none of the history runs the final gate.

I used it for the To Do rebuild and the Free My TV restyle. The plan above is this site's own.

/do is a Claude Code skill, a set of written instructions that Claude follows, and I built it with Claude Code.

I designed the loop and its checks. The To Do rebuild was its first real test.

I've since made 2 versions. /do-light is for small jobs: Claude drafts, one fresh critic checks the draft, and Claude fixes what it finds.

/do-sonnet runs the builders on Sonnet, a cheaper model, to see if the quality holds. It built this site.