Make Your Claude Chat Last Twice as Long
Subagents: let Claude send a helper to do the heavy reading
If You Only Read One Thing
Paste this into Claude. It adds one rule to your CLAUDE.md so that, from now on, Claude offloads the heavy reading to a subagent automatically — keeping your main chat light without you having to remember. The rest of this page explains what that's doing and why it works.
The Problem
There's no worse feeling than a Claude chat that's been going for ages — it starts slowing down, straining, getting things wrong. And you know what's coming: opening a new chat and having to bring it back up to speed with everything you've already explained.
Starting fresh is good advice, and eventually you'll always need to (that's what the context window is all about). But there's a way to prolong it — to get a lot more out of a single chat before you have to.
The Fix: Subagents
Right now it's just you and one Claude. Every bit of research, every file it opens, every tangent — it all piles into the same memory. That's why it strains: it's holding too much at once.
But Claude can manage its own helpers. This is built into Claude Code, and Anthropic calls them subagents. So instead of Claude doing a big side-quest itself — reading ten files, filling its own memory up — it sends a subagent off to do the digging in a separate memory. The subagent does all the heavy reading, then comes back and tells Claude just what it found.
Without subagents
One box holding everything. It strains.
With subagents
Helpers read in their own memory and report back just the answer.
Claude gets the answer without ever holding the mess. Your main chat stays light and sharp — so you're not starting over every twenty minutes.
Three Ways To Use One
It's catching on — Claude sometimes does it on its own now. But you can also tell it to:
Just ask
In plain language: "use a subagent to dig through these files and tell me where the auth logic lives." Claude spins one up, it reads, it reports back.
Name a specific one
If you've got custom subagents set up, point at one directly — "use the code-reviewer subagent on the auth changes." Naming it makes delegation reliable.
Make it automatic (my favourite)
Add one line to your CLAUDE.md and Claude delegates the heavy stuff on its own, every time — no need to remember. The exact line is below.
The One Line For Your CLAUDE.md
The prompt at the top of this page is the easiest way — Claude edits the file for you. Prefer to paste it in by hand? Here's the exact rule it adds. Drop it into your CLAUDE.md (project root, or ~/.claude/CLAUDE.md for every project):
One honest note: CLAUDE.md is a strong nudge, not an iron rule — Claude weighs it against the task in front of it, so it won't delegate every time. Keeping the instruction specific (like above) makes it stick far better than a vague "be efficient."
When To Use One
Good for
- →Reading through a pile of files to find one thing
- →Searching the whole codebase for where something lives
- →Digging through logs, docs, or research to answer a question
- →Any tangent that means a lot of reading you won't need again
Not worth it for
- ·The actual work you want in your main chat's memory
- ·A quick one-file edit — delegating it is just overhead
- ·Anything where you want to see every step as it happens
What It Actually Saves You
The real win is that your main chat stays lean, so it lasts far longer before it starts to strain — you're not re-briefing a fresh chat every hour. There's a usage upside too, but be clear-eyed about it: the subagent still has to read those files, so the reading isn't free. What you avoid is that pile of files sitting in your main chat and getting re-read on every single reply after that. Over a long session, keeping it out of the main thread is what adds up.