The short answer: treat Claude Code as a small team, not a chat window. Split the work into jobs that don't touch the same files, give each job to its own subagent, keep one shared rulebook (CLAUDE.md), and do the final check yourself. That is how I rebuilt most of this website in an afternoon.
This way of working is everywhere right now. Boris Cherny, who created Claude Code at Anthropic, shared his setup earlier this year: several Claude sessions running at once, a CLAUDE.md where every mistake becomes a rule, and a loop where the agent tests its own work in a browser. VentureBeat wrote it up and developers have been copying it since. Here is what it looks like on a real project, including the part that went wrong.
What does "running Claude Code like a team" mean?
It means one main session acts as the lead and hands out work. The lead reads the codebase, writes a short brief, and starts subagents in the background. Each subagent gets one job, a list of the files it owns, and the rules. The lead doesn't write the pages itself; it reviews, builds and ships.
On theaikhan.com I needed five things at once: the blog and guides pages, the FAQ and About pages, the projects pages, and a site-wide SEO pass. Four subagents, four jobs, running in parallel.
Why file ownership matters more than the prompt
The one rule that made it work: every agent gets its own files, and may not edit anyone else's. Two agents editing the same stylesheet at the same time is how you lose an hour.
So each brief said, in plain words:
- the files you own (for example
app/(site)/faq/page.tsxand a newcomponents/faq/folder) - the shared contract you may rely on, but not change (the page header and its props)
- what to report back if you need a change outside your files
When the projects agent finished, it reported that it had built its own breadcrumb instead of waiting for the shared one. That's the system working: no collisions, and a note I could act on.
One brief, written once
Instead of repeating the rules in four prompts, I wrote them into one brief file every agent read first:
- use the design tokens only, no raw colours
- never invent a number; a missing metric shows "—" and "Metric to confirm"
- one
h1per page, answer-first copy, structured data on every page - don't run a full build (the disk was nearly full) — type-check and lint your own files instead
That last line is the kind of rule that only exists because something broke before. That's what CLAUDE.md is for: the project's memory of its own mistakes.
What went wrong
Two of the four agents hit the usage limit halfway through. They stopped mid-task.
Because each one owned its own files, nothing was half-merged into someone else's work. I checked what each had left, then started smaller, cheaper agents with a narrower brief: "finish and check this, here's what already exists." They completed the work.
The lesson: pick the model per job. Big design work went to a stronger model; finishing and checking went to a faster one. Claude Code now lets you pin the model for every subagent in a workflow (it shipped in version 2.1.296 on 9 October, per Tech Insider), which makes this a setting instead of a habit.
Who checks the work?
You do, every time. After the agents finished I ran the type-checker, the linter and a production build, then loaded every page at desktop and phone widths to check for sideways scrolling, console errors and missing structured data. I found and fixed two things myself (page titles in the wrong order, one description that was too long).
Agents are fast and confident. The verification step is where the quality comes from.
The setup in five lines
- One lead session that plans, reviews and ships.
- One brief file with the rules, read by every agent.
- One job per subagent, with the files it owns written down.
- The right model per job, and a plan for when one stops.
- You run the build and look at the result before anything is merged.
If you'd rather have a two-agent loop, where one model plans and checks and the other writes the code, I wrote a step-by-step guide for that: Claude x Codex Loop.
Have a problem you want built this way? Open a case file.