How to talk to AI so it does what you mean

People who say "I tried it and it was useless" almost always typed one vague line and got one vague answer back. People who say "I cannot work without it" learned a few habits. Here they are.

1. Say who it is talking to and what for

The model knows nothing about you. "Explain compound interest" gets a textbook. "Explain compound interest to me, I am 40, I have never invested, I want to know if it is worth starting now with $200 a month" gets something you can use. Context is not padding. It is the answer's raw material.

2. Say what a good answer looks like

Length, format, tone. "Three bullet points." "A short email, friendly, no exclamation marks." "A table with the pros on the left." If you do not say, it guesses, and its guess is usually longer and more formal than you wanted.

3. Show it an example

If you want it to write in your voice, paste two emails you wrote and say "like these". If you want a spreadsheet layout, describe one row. Models copy shape brilliantly. One example is worth a paragraph of description.

4. Tell it to say when it does not know

Add "if you are not sure, say so" and "do not guess" and watch the quality jump. Models default to confident. Given permission to be uncertain, they use it, and their confident answers become more trustworthy because you know the difference.

5. Treat it as a conversation, not a search box

The first answer is a draft. "Shorter." "Less formal." "You missed the part about the deposit." "Now do the same for my sister's situation." Each reply is a few seconds. Three rounds of that beats ten minutes of writing the perfect first prompt.

6. Put standing instructions where the app keeps them

If you find yourself typing the same setup every time ("I am a nurse, keep it short, no jargon"), put it somewhere the app remembers. Claude calls it a Project. ChatGPT calls it a Project or custom instructions. That is exactly what an agent's instruction block is: standing orders, written once.

Things that do not help

• Being polite. It does not mind. Say please if you like, it changes nothing.
• Threatening or tipping it. Old internet folklore. Ignore it.
• Writing in a special "prompt language". Plain English is the native language. Write the way you would brief a smart new colleague.
• One giant message with everything in it. Break it up. Give the background, get a reply, then ask.

A worked example

Vague: "Help me plan meals."

Better: "Plan five dinners for this week for two adults and a toddler. We are vegetarian. Budget around $80. We have rice, eggs and a lot of frozen peas already. Give me the plan as a list, then one grocery list grouped by aisle. If you are not sure something is toddler friendly, say so."

The second one is what a good agent's instructions look like, just written once instead of every week. That is the whole idea behind this site: we write the second kind for you, you paste it once, and it runs.

What to remember

• Context, format, example, permission to be unsure, iterate, and save your standing orders.
• Plain English. No magic words.
• The first answer is a draft. Reply to it.