3 AI helper role files: researcher, writer, checker
Here are the three helper role files you asked for. Each one is a short text file with one job. In Claude Code, save them in .claude/agents/. In other tools, paste the role at the top of a fresh chat or into a custom assistant.
researcher.md
---
name: researcher
description: Looks into one question on the web and returns a short, sourced summary. Run several in parallel for options.
---
You research one question. Read CLAUDE.md first.
- Prefer official pages and primary sources. Give the URL for every fact.
- Mark anything you couldn't verify as "unverified". Never invent numbers.
- Don't log in anywhere, don't fill in forms, don't download files.
Return: at most 150 words — the answer, the 3 most important facts with sources, and what's still unknown.
writer.md
---
name: writer
description: Writes a first draft (email, post, page) in my voice from a short brief.
---
You write first drafts. Read CLAUDE.md and memory/INDEX.md first; follow my tone notes.
- Ask for nothing; if the brief is missing something, write [brackets] where it goes.
- Keep it shorter than you think. Lead with the point.
- No claims or numbers that aren't in the brief.
Return: the draft, then one line listing anything you put in [brackets].
checker.md
---
name: checker
description: Checks a draft against my rules before I send it. Use for emails, posts and documents.
---
You check drafts. Read CLAUDE.md first.
For each draft:
1. Does it follow "How to answer" and "Never"?
2. Are all numbers and names backed by a source in the draft?
3. Anything that sounds like a promise we can't keep?
Reply with: OK, or a list of fixes (one line each). Don't rewrite the draft.
Want the whole system?
Build Your Own AI Assistant System · $5 — about a coffee
The rules file, memory, sub-agents and automations — with copy-paste templates. Bonus: the Note tool (Windows) that makes your AI fix only what you point at.
What a helper is, and when it pays off
A helper is a separate assistant with one job, which reports back a short result. One assistant doing everything in one long chat gets slow and muddled. Splitting the work fixes that in three situations:
- Parallel research. Three helpers each look at one option; you get three short summaries instead of one tangled thread.
- Write, then check. The checker hasn't seen the writer's reasoning, so it spots mistakes the writer talked itself past.
- Repetitive work. The same task for ten items, each handled cleanly.
If a job is a quick lookup, don't bother. Ask the main assistant.
Brief a helper like a new colleague
A helper knows only what you tell it. Use this template:
Goal: [one sentence]
Already known / ruled out: [so it doesn't repeat your work]
Where to look: [files, sites]
Return: [what, and how long, e.g. "a 150-word summary"]
Example:
Goal: Find three project-management tools under 20 euros per user per month.
Already known / ruled out: I've already tried Trello and dislike it.
Where to look: each tool's official pricing page.
Return: a 150-word comparison with the pricing page link for each.
The most common failure is a one-line brief like "research project tools". The helper guesses, and you get an answer to a question you didn't ask.
Keep the cost down
Helpers use your credits or usage limit quickly, because each one starts a conversation of its own. Three rules:
- Use helpers for parallel or independent work only. Not for something you could do in one message.
- Pick a smaller, cheaper model for research and checking. Keep the strongest model for decisions and final writing.
- Cap the number. Add this line to your rules file: "At most 3 helpers at a time; tell me before using more."
Common mistakes
- Giving a helper your whole life. A role file should be ten lines, not a profile. Point it to CLAUDE.md for the rest.
- Letting the checker rewrite. It should list fixes, not produce a second draft. Otherwise you can't see what changed.
- Trusting research without sources. The researcher file demands a URL for every fact. Open one or two before relying on them.
- Giving helpers permission to send or buy. Helpers return results. You decide what happens next.
Test it
Take an email you'd normally send. Ask the writer for a draft from a three-line brief, then pass the draft to the checker. If the checker finds nothing, try a draft with an invented number in it and confirm it gets flagged.
Helpers are one of five parts of the system. The rules file, memory notes, project folders and the morning brief are in the full guide, with every template ready to copy.
Want the whole system?
Build Your Own AI Assistant System · $5 — about a coffee
The rules file, memory, sub-agents and automations — with copy-paste templates. Bonus: the Note tool (Windows) that makes your AI fix only what you point at.