Skip to content

What Do I Actually Type Into It?

November 11, 2026

You’ve heard ChatGPT will save you hours. You open it, type what seems obvious, and get back a page of waffle. So you wonder what everyone else knows that you don’t.

Probably nothing. Most of what’s sold as the secret is not a secret, and some of it doesn’t work.

This post covers what to put in the box, what to leave out, and why the first answer is rarely the one you use.

Why you get waffle

When you ask a general question, you get the most likely answer to a general question. That’s how these tools work. They produce what usually comes next.

The tool is missing the things only you know. The gap in your schedule next week. The price you really charge. Who the customer is and why this job is awkward.

Two earlier posts cover the biggest pieces of this, so we’ll point to them instead of repeating them:

If you only do those two things, you’ll be ahead of most people. The rest of this post is what’s left.

Four things to put in every request

1. Who it’s for and what you want to happen

“Write a follow-up email” gives it nothing to aim at. Try this instead: “Write a follow-up to a customer who got a quote nine days ago and has gone quiet. I want her to reply this week.”

The outcome you want is part of the instruction. Without it, the tool aims at the middle.

2. How long, and what shape

Say it out loud. Four sentences. A checklist of five items. A table with three columns.

If you don’t, you tend to get an essay, because an essay is the average answer.

3. What to leave out

This one gets skipped. “No exclamation marks. Don’t use the word ‘elevate.’ Don’t promise a discount. Don’t invent details I haven’t given you.”

A list of things to avoid takes ten seconds to type and cuts a lot of cleanup.

4. Why you want it

Give the reason, not only the order. “Keep it short because she reads these on her phone between jobs” tells the tool more than “keep it short.”

Anthropic, the company that makes Claude, publishes advice on this. Its guidance tells people to treat the tool as a brilliant new employee who lacks context on how you work, and says the more precisely you explain what you want, the better the result. It also offers a test you can use: show your request to a colleague who knows nothing about the job and ask them to follow it. If they’d be confused, the tool will be too. (That’s Anthropic’s own documentation. It’s the maker’s advice, not independent proof. It’s still a good test.)

What that looks like on a job that isn’t writing

The example below is made up. It’s an illustration, not a customer story.

Say three emails arrived before 9 and you’re on site until 2. You only have time for two replies.

The vague request: “Help me with my emails.”

You’ll get an offer to help, and a list of questions back.

The better request:

Here are three customer emails that came in this morning. I’m on site until 2pm and can answer two of them before then. Tell me which two to answer first, in what order, and give me one line on why for each. Don’t draft any replies yet. I need to decide first.

[the three emails pasted below]

Now the tool has a limit (two of three), a deadline (2pm), a format (one line each) and a boundary (no drafts). The answer might say that the customer disputing a charge goes first, because money disagreements get worse the longer they sit, and that the one asking whether Thursday still works goes second, because Thursday is close. The general pricing question can wait.

You can check that in under a minute. You know these customers and can overrule it. The tool sorted. You made the call.

Expect two or three tries

Treat the first answer as a starting position.

Then you steer it. “Shorter.” “Lose the first paragraph.” “Put the date in the first line.” “Make the second option cheaper.”

If you stop after one try, you’ve only seen the draft. The second and third messages are part of the process.

The tricks you’ve been sold

If you’ve searched for this topic, you’ve probably seen prompt packs and “secret phrases.” Here’s what the evidence says about two of the common ones.

“Act as a world-class marketing expert.” Researchers tested this. A 2024 paper from the Findings of ACL: EMNLP tried 162 different roles on 2,410 factual questions across four families of AI models. Adding a role to the prompt did not improve performance compared with adding none. Picking the best role for each question would have helped, but nobody could pick it ahead of time any better than chance. (Zheng and others, ACL Anthology) Note that this tested factual questions, so it says nothing about whether a role changes the tone of a piece of writing.

Special structures that expire. A paper posted to arXiv in August 2026 tested five prompting techniques on code generation, using 218 Python functions, six models and three pairs of older and newer versions. The results were mixed across model families. The benefit faded for the GPT models, stayed substantial for Qwen, and was mixed for Mistral. The authors conclude that prompting strategies need adapting for each model family. (Rudyk, Oertel and Hebig, arXiv) It tested code, not everyday writing, and we haven’t confirmed whether it has been through peer review. arXiv lists the venue as ICSME 2026. Treat it as a hint, not a final answer.

Our reading: a prompt pack may be worth less each time the tools update. The part that keeps its value is giving the tool your own facts. That never goes out of date, because the tool can’t get your facts anywhere else.

You don’t need a pack. You need the four things above.

One warning about better prompts

A better request does not fix a task the tool is bad at.

A field experiment with 758 consultants at Boston Consulting Group found that AI helped on some tasks and hurt on others. On a task outside what the tool does well, people using it got less accuracy than people working alone. According to the summary from the Harvard Business School AI Institute, a group that had been given extra prompting training did worse on that task, not better. (HBS AI Institute; the paper is in Organization Science, March 2026.) We read the summary, not the paper.

So a good prompt makes a good task better. It doesn’t turn a bad one into a good one, and it can make you more confident while you’re wrong.

Pick tasks where you can check the answer yourself. Sorting emails, pulling facts from a document you pasted, and drafting something you’ll read before it goes out all qualify. Tax rules and legal wording do not. For the checking side, see How do I know the AI isn’t making it up?.

Where this leaves you

You weren’t doing something wrong. You were asking questions with no business in them, and the tool answered the way it always does.

Next time, say who it’s for and what should happen. Give it the length and the shape. Tell it what to avoid and why. Paste in your own material. Then answer back when the first try is off.

The short version: the tool doesn’t know your business. Tell it the parts that matter, and tell it again if the first answer misses.

Keep reading

Want to talk it through?

Tell me about the task that takes up your team's time.

Newsletter

Plain-language notes on software and AI for small business

One short email when there is something worth sharing. No spam. Unsubscribe any time.