How to talk to your agent so it gets things right

Most disappointing answers from an agent come from vague requests, not a weak model. You don't need "prompt engineering". You need what you'd give a capable new assistant: the goal, the context, the limits, and what a finished result looks like.

The four-part brief

1. Goal: what you want done.
2. Context: what the agent should know first.
3. Limits: length, tone, budget, deadline, things to avoid.
4. Finish line: the format you want back, and where to put it.

The second one gets something usable on the first try.

Set boundaries before action

Your agent can do things, not just write things: create files, run code, use your connected apps. Tell it what it can do on its own and what needs your OK.

• Researching, drafting and organising: fine to do directly.
• Sending an email, publishing a post, deleting files, changing an account or spending money: ask first.

Put this in a sentence at the start of a task ("draft it, don't send it") or permanently in its personality file. See give your agent a personality and approvals, privacy and safety.

When it says "done" but nothing happened

This is the most common trust-breaker. Polished wording isn't proof that an action ran. Usually one of these happened:

• It described a plan, and you both read it as finished.
• The action was waiting for your approval and never ran.
• It worked on a different file or folder than you expected.
• It answered from memory without checking the current state.

The fix is to ask for evidence every time it matters:

Then check it yourself, for example in Files.

When it changes too much

If you asked for one fix and it rewrote half the file, tighten the scope:

• Name the exact file and the exact part to change.
• Say what must stay the same.
• Ask for a short summary of what it plans to change before it changes anything.

"Only change the second paragraph of pitch.md. Leave everything else exactly as it is. Tell me what you'll change before you do it."

When it keeps asking the same question

If it asks something you've already answered, the answer probably only lived in an old chat. Ask it to save the fact to memory: "Remember that my business is called Brightbooks and my customers are freelancers. Save it to your memory file." See how your agent's memory works.

When it replies in the wrong language

Agents often mirror the language of the last message, a pasted document or a web page. Tell it once, then make it permanent: "Always reply in English, even if I paste text in another language. Add this to your SOUL.md."

Keep jobs small and check the result

Big, open-ended requests ("grow my business") produce big, vague answers. Break the work into steps with a result you can inspect, finish one, then ask for the next. Long-running conversations also get worse over time; see long chats and starting fresh.

Frequently asked questions

Why does my agent give generic answers?

It usually doesn't have enough context. Add who it's for, the constraints and the output format you want. Saving key facts about you to memory helps every future answer.

How do I stop my agent from doing things without asking?

Tell it which actions need your approval, like sending, publishing, deleting or spending, and ask it to add that rule to its personality file so it sticks.

My agent edited the wrong file. Can I undo it?

Ask it to show what changed and restore the earlier version if it can. For anything important, keep it in GitHub so every change can be undone; see use GitHub with your agent.

Should I use a higher effort level?

Only for hard problems. Higher effort makes the model think longer, which is slower and uses more of your monthly AI allowance.