Free tool

Build a Claude helper that takes over a job

Claude can hand a whole kind of work to a specialist you set up once, called a subagent. A reviewer that only reads, a helper that writes your release notes, a researcher that goes and looks things up: each one works in its own space and hands back just the result. This builds you that helper as a file, from a few plain answers.

You do not need to know how to code to use this. The bottom of the page tells you where the file goes, including a way that involves no typing at all.

Name it

A short name so you and Claude can refer to it. Lowercase words with hyphens work best, and this tidies whatever you type into that shape for you.

What should this helper be called? Something plain, like code reviewer or release notes writer.

When Claude should hand it the work

This is the one that matters most. Claude reads this line to decide when to pass a job to the helper instead of doing it itself, so describe the situation, not just the subject. "When reviewing code before it goes out" works far better than "code review".

When should Claude use this helper? Say the kind of moment it should step in.

What it should do

The instructions the helper follows every time it runs. Write them the way you would tell a new person how to do this one job well: what to look for, what to leave alone, how to hand back what it found.

Write the helper's instructions. Plain sentences, as long or short as the job needs.

What it is allowed to do

Tick only what this helper needs. Fewer ticks is safer: a helper that can read but not change your files cannot break anything, which is exactly what you want for one that reviews or researches. Leave them all unticked and it can do anything you can.

Read and search my files. Look through the project to understand it, without changing a thing. This on its own is enough for a helper that reviews or explains. Create and change files. Let it do the actual writing and editing. Tick this only when the helper is meant to make changes, not just look. Run commands. Let it run things like your checks or scripts. Handy for a helper that tests or builds, worth leaving off otherwise. Look things up online. For a helper that needs current information from the web. Leave it off if the work is private.

How much thinking power it needs

Simple, repeated jobs run fine and cheap on a smaller model. Genuinely hard reasoning is worth a bigger one. If you are not sure, leave it on the first option.

The same one I am already using. A safe default. It matches whatever you have set for your main session. Fast and cheap. Best for simple, high-volume jobs where speed and cost matter more than depth. Balanced. A sound middle for most real work. Most capable. For the genuinely hard jobs, at a higher cost.

Your helper file

This updates as you answer. Nothing here leaves your browser.

Where the file goes

The easy way, with no typing: hit Copy, then paste the whole thing to Claude and say save this as a new helper I can use on every project. It knows where these files live and will put it in the right place.

By hand: it is a small text file whose name ends in .md, and it lives in a folder called agents inside the .claude folder. For one project, that is .claude/agents in that project. To use the helper everywhere, put it in .claude/agents in your home folder instead. Name the file after the helper, like code-reviewer.md.

How to tell it worked

Start Claude fresh and ask it to do the kind of job you described. It should say it is handing the work to your helper by name. You can also just ask Claude to use that helper directly.

Start it read-only if you are unsure. A helper that can look but not touch is the safest way to try one out, and you can hand it more later by ticking another box and saving the file again.

Build a Claude helper that takes over a job

Your helper file

Where the file goes

How to tell it worked

Get the next one