← All use cases

中文

Automation & agents · Claude

Run a one-week trial where three agents handle research, content and support

Idea + starter prompt, not run by us yet

For: Solo owners who want to test how much a small agent team can safely cover

A solo owner shut their inbox for seven days and ran three narrow Claude agents. One researched industry news, one wrote posts and emails in their past voice, one handled messages, FAQs and call booking. The week produced 14 pieces, 4 discovery calls and about 22 hours saved. It also produced hollow copy, an invented statistic and a canned reply to a vulnerable message.

Steps

  1. Start with three narrow jobs rather than one agent that does everything.
  2. Give each agent a short written rule on what it must hand to you: money, refunds, upset customers.
  3. Check every factual claim before anything is published. A made-up statistic went live within minutes in this trial.
  4. Read the support replies daily, especially to anyone who sounds stressed or upset.
  5. After a week, decide which jobs to keep with agents and which to take back.

Starter prompt

Starter prompt. Change the [brackets] first.

Trial first: before the real week, run each helper on made-up material for one day: [5] invented customer messages, [3] sample news items and one pretend post. Save everything in [trial-week]. Do not touch my real inbox yet.

I run a one-person business that does [what you sell]. For the trial week I want three narrow helpers. Research collects industry news in [topic]. Writing drafts posts and emails in my style from [sample files]. Support answers FAQs and offers booking times. Show me the plan and the rules each helper follows. Write down when each must stop and ask me, including anything about money or a customer who seems upset.

Everything is a draft. Stop and wait for my yes before any message is sent, any post goes live or any call is booked. Every statistic needs a source link.

Original post: Medium · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.