The 3 Prompts to Hand a Big Job to a Team in Claude Code
Most people make one Claude chat do everything, then it runs out of context halfway through. These 3 prompts hand the big job to a team of sub agents, keep your main chat clean, and make the next job faster.
Here's how most people use Claude on a big job. One chat, doing every piece, one after another.
Halfway through, the chat is so full it starts forgetting things, or it runs out of context.
Claude Code can run a team instead. There's nothing to install. You give Claude one big job, and it hands the pieces to sub agents. Each one works in its own space and brings back only a short summary. So your main chat stays clean, and you can keep chatting with it.
When I add a new course to my library, every part goes to its own sub agent at the same time.
Before you start
Open the Claude desktop app, switch from Chat to Code at the top, and pick the folder with the files for your job.
Sub agents pay off on big jobs with pieces that don't depend on each other: 3 course modules, 5 competitor pages, 10 reel transcripts. For a small, quick change, just ask normally. A team is slower and uses more of your Claude usage on tiny jobs.
Prompt 1. Hand the big job to a team
Fill in your job, then paste it.
Use sub agents to do this in parallel: [your big job, like "turn each of the 3 module files in this folder into a lesson summary and a 5 question quiz"]. 1. Split the job into pieces that don't depend on each other. Tell me the split in a short list before you start. 2. Give each piece to its own sub agent, running in the background at the same time. 3. Give each sub agent only what it needs: the file names to read and exactly what to hand back. 4. When they're all done, put the results together and show me what each one did in one line each. Use only what's in my files. Never make up facts, numbers or quotes. If a piece needs something that isn't there, say so.
Prompt 2. Keep your main chat clean
Add this to the end of Prompt 1 on a job with big results, like long documents. It stops the results from filling up your main chat.
Each sub agent saves its full result as a file in a folder called results, then hands back only the file name and a 3 line summary. Don't paste the full results into this chat.
Now your main chat holds a few lines per piece instead of every page. You can keep going for hours.
Prompt 3. The second track: make the next job faster
This is the part most people skip. After every big job, paste this:
The job is done. Now the second track: how to make the next one faster. 1. Write down what slowed this job down, what you'd split differently, and what to do first next time. Keep it to 5 short lines. 2. Add them to a file called LESSONS.md in this folder, under today's date. Don't delete older lessons. 3. If there's no CLAUDE.md in this folder, make one. Add this line to it: "Before any big job, read LESSONS.md first. After any big job, add what to do faster."
Claude Code reads CLAUDE.md at the start of every chat in this folder. So next time, it reads your lessons before it starts, without you asking.
The limitations you should know
- A team only helps on big jobs with separate pieces. On small or back-and-forth jobs, one chat is faster.
- More sub agents use more of your Claude usage. Start with 3 to 5.
- Sub agents don't see your chat. Each one only knows what Prompt 1 tells it, so be clear about the files and what to hand back.
- Read the results. The team does the work. You still decide what's good.
- Technical and want more? Matt Pocock's free chief-of-staff skill does a bigger version of this for coding projects: https://github.com/mattpocock/skills
Want the next guide? Follow @iamzachcrawford