Technique

Put Claude's Juniors to Work

Two lines that make delegation standing behaviour, and the ones you type when you only want it for this job

Start Here: Paste This

Here from the juniors video? This is the line. Paste the prompt into Claude Code exactly as written — nothing to fill in. Claude adds the staffing rule to this project's CLAUDE.md, shows you what it wrote, and names the first job it would now delegate. From then on, small mechanical work goes to juniors on smaller, cheaper models and the big brain sticks to the thinking.

Paste this into Claude

Please add a staffing rule to this project's CLAUDE.md so you delegate small jobs to cheaper models automatically. Open the CLAUDE.md in the project root — create it if this project doesn't have one yet. If there's already a Delegation section, merge this rule into it rather than duplicating: ## Delegation For small mechanical jobs — renames, simple lookups, one-line fixes, formatting — hand the work to a subagent on a smaller, cheaper model instead of doing it yourself, and review its work before it lands. Save your own effort for the thinking: planning, judgment, and anything ambiguous. When you're done, show me exactly what you added or changed, and tell me the first job you'd now delegate in this project so I can see it working.

Not ready to change any files? The lines further down do the same thing one job at a time, typed straight into the chat.

Never heard of CLAUDE.md? It's the document Claude reads before every job — the prompt above creates it if you don't have one, and the full guide to it is here.

What's Actually Happening

By default it's one Claude doing every part of a job personally — the thinking and the grunt work, all on the biggest model you've got, all piling into one memory. That's the senior partner renaming files, and it's why long chats strain and limits vanish.

But Claude can manage helpers — Anthropic calls them subagents. It hands a piece of work to a junior copy running in its own separate memory, on a smaller model where the job allows, and gets back just the result. The partner decides, a junior does it, the partner checks it after — you can simply ask for one in plain language, or use the lines on this page to make it standing behaviour.

You'll know it's working when a little task spins off to the side instead of Claude opening file after file in your chat.

The Two Lines

The staffing line — for your bill

Small mechanical jobs go to juniors on cheaper models, reviewed before they land. Tokens are the new billable hours — as AI takes on more of your work, the gap between one expensive brain doing everything and a staffed team is real money. It's the prompt at the top of this page.

The reading line — for your chat

Heavy reading — piles of files, logs, research — happens in a junior's memory instead of your conversation, so your chat stays lean and lasts far longer before it strains (that strain is the context window filling up). This prompt adds it:

Paste this into Claude

Please add a delegation rule to this project's CLAUDE.md so you lean on subagents for heavy reading automatically. Open the CLAUDE.md in the project root — create it if this project doesn't have one yet. If there's already a Delegation section, merge this rule into it rather than duplicating: ## Delegation For heavy, exploratory work — reading lots of files, searching the codebase, digging through logs or docs — use a subagent instead of doing it in the main thread. Let it do the reading in its own context and report back just what it found, so the main conversation stays lean. When you're done, show me exactly what you added or changed.

Or Just Ask For One, This Once

You don't need a rule in a file to get a junior. Ask for one in the moment and Claude spins it up for the job in front of you, does the work off to the side, and comes back with the result. The word subagent is doing all the work in these lines. Everything else is just the job, so swap it for whatever is actually in front of you.

When you want it checked

Review the changes on this branch with a subagent, then tell me what it found and what you would fix first.

When it needs to read a lot first

Before you change anything, send a subagent to read through this project and report back on how it works. Keep the reading in its context rather than ours.

When there are several jobs at once

Split this across a few subagents, one job each, and bring their findings back together. They are only reading, so put each one on a smaller model.

Two things worth adding whenever you ask. Name a smaller model if the job is mechanical, because Claude will otherwise staff it with the expensive one. And say what you want back, because the report a junior hands over is the only thing that survives. Everything it read and didn't write down goes when it finishes, which is the point, and also the catch.

Using The Claude App, Not Claude Code?

The feature lives in Claude Code, but the idea works anywhere: you play the junior's memory yourself. When a big side-quest comes up, do it in a separate chat and paste just the conclusion back into your main one — the heavy reading lives and dies in the side chat, and your main conversation only ever carries the answer.

The Honest Bit

Juniors still use tokens — the reading isn't free, and delegating a quick one-file edit is just overhead. The staffing line saves the premium you stop paying for the big model's time on jobs that never needed it; the reading line saves your chat from re-reading a pile of files on every reply. And CLAUDE.md is a strong nudge, not an iron rule — specific instructions like these stick far better than a vague "be efficient".

Want to get better at this?

Let's set up your agent ecosystem.

This page is one job. The whole thing is a set of them running off the same foundation, the folder your AI works out of, the brief on how you like things done, the accounts it can reach and the skills that run off all of it, so your admin gets done the way you would do it. It is what I teach one to one, on your own machine and against your own work, and the quickest way to find out what yours would look like is a proper chat about it.

Book a call Or read about the course first

Not ready for a call? Start with the free agent series, what an AI agent actually is, and build up from there.