AI Guides / Learn

How to talk to an AI agent: your first hour

Tell an AI agent the result you need, the source files, the limits, and the final check. Start in a safe working folder, review every change, and keep the first task small.

Pick one useful outcome

Don't begin with a long list of clicks. State what the finished work should look like. The agent can choose the steps inside the limits you set.

A weak request says: "Open column C, copy each value, and sort it."

A stronger request says: "Read inventory.csv. Create low-stock.csv with every item below 10 units. Keep the product ID and supplier. Confirm the new row count when done."

The stronger request names the input, output, rule, and check. It gives the AI agent room to work without hiding the finish line.

Choose a task you understand well. Good first tasks include sorting a copied list, drafting replies, or comparing two exports. Avoid live systems and automatic sending during your first hour.

Make a safe working folder

Create one folder for the task. Put only the files the agent needs inside it. Don't open the agent in your whole desktop or home folder.

Work with copies, not your only records. Keep raw files outside the folder or mark them read-only. Ask the agent to create a new output file instead of replacing an input.

Use plain names that explain each file. customers-original.csv is clearer than final2.csv. Tell the agent which file is the source of truth.

Folder limits reduce the damage from a wrong instruction. They also make the file diff easier to review.

Give examples of your own work

An LLM doesn't know your unwritten style. Give it two or three examples you already approve.

For customer replies, include a few good past replies. For a weekly report, include last week's approved report. Remove private details that the task doesn't need.

You can add a short AGENTS.md or CLAUDE.md rules file. Keep it between 200 and 500 words. Include facts that would prevent a real mistake.

Useful rules might say:

• Never delete or replace an original file
• Save every result in the output folder
• Use only facts found in the supplied product sheet
• Round money to two decimal places
• Report any row you couldn't process

Skip general advice the agent already knows. Short, testable rules are easier to follow.

Ask for proof, not a promise

"Done" isn't proof. Ask the agent to check the result against facts you can inspect.

For a spreadsheet, compare input and output row counts. Spot-check five normal rows and several odd rows. Look closely at blank fields, unusual symbols, and missing email addresses.

For changed documents, look at the diff. A diff shows what was added, removed, or replaced. Confirm that only the intended files changed.

For drafts, check names, dates, prices, links, and claims. Don't let the agent send or publish the work during this first task.

The agent harness may run its own checks. Your review still matters. A check can pass while the original instruction was wrong.

Avoid five beginner mistakes

1. Don't pack a whole project into one request. Split large work into stages. Ask for an inspection and plan first. Review it before the agent acts.
2. Don't use a vague chatbot prompt. Give the input, desired result, limits, and check. The agent can't read your mind.
3. Don't remove permission limits. Avoid unrestricted modes around important data. Use a duplicate folder and keep approval prompts on.
4. Don't accept work without a check. Name the row count, file list, length limit, or calculation that must be verified.
5. Don't mix unrelated work in one long chat. Start a fresh session when the task changes. Old context can confuse the next job and raise usage costs.

These habits matter more than clever prompt wording. Clear boundaries make a useful result easier to review.

Use your first hour well

Spend the first few minutes choosing a small task and copying its files. Then write the outcome, limits, and check in plain language.

Let the agent inspect the folder before it changes anything. Review its plan. Ask questions if it names a file or action you didn't expect.

After it works, read the summary and inspect the diff. Open the output yourself. Spot-check the content, then decide whether to keep it.

Write down one rule the agent needed but didn't have. Add that rule before the next similar task. This turns each run into a clearer process.

If you're asking "how to use AI in my business," don't chase a perfect prompt. Build one task you can explain and verify.

Common questions

Should I tell the agent every step?

No. State the outcome, inputs, limits, and check. Add a required step only when your process truly depends on it.

How much business context should I share?

Share only what the task needs. Use a short rules file and a few approved examples instead of your full business history.

When should I stop the agent?

Stop when it reaches outside the agreed folder, asks for risky access, or changes work you didn't name. Review before continuing.

Next step

Type what your business does and see three examples, or send us a message about one task you'd like help with.

Keep reading

• Hermes Agent for business owners: what it is, what it costs, how to start
• What is an AI agent?
• What is an LLM?
• What is an agent harness?
• Codex vs Claude Code vs Hermes Agent: which should a business owner start with?