What should I hand to AI, and how do I ask for it?
Hand AI the work you can describe and check, and keep the work you can’t. A prompt is a brief: what you want, who it’s for, the material it should use, and what a good result looks like. The wording matters far less than whether those things are in it. This page is the version of that lesson we send client teams. One decision tree, one six-line brief, one before-and-after, and a plain table of which tool fits which job. No scorer and no sign-up.
Should this go to AI at all?
Three questions and four places you can land, and one of them is “do it yourself.” That outcome is on the chart on purpose. If you can’t describe a good result, you can’t check for one, and an unchecked answer is worse than no answer.
Work the problem by hand once. Come back when you can describe the target.
Paste the notes, the call summary, the spreadsheet. Without them it will invent the gap.
Ask for options and the reasoning. You pick.
Draft, summary, first pass. Read it before it leaves your hands.
The six-line brief every good prompt contains.
Most weak prompts aren’t badly worded. They’re missing lines. Before you send one, check it against these six, and if a line is blank and you don’t know the answer, that’s the work, not the typing.
What you want
The deliverable, named. A draft email. A list of five options. A one-paragraph summary. “Help me with” is not a deliverable.
Who it's for
The reader, and what they already know. A customer who hasn’t replied in two weeks reads differently than a new hire on day one.
What it must not do
The lines it can’t cross. No discount. Don’t mention the delay. Under 120 words. These are the constraints a person would never forget and a model will.
The shape
Bullets or prose. A table. Three versions to choose from. Say it, or you get the model’s default shape, which is rarely yours.
Your material
Paste what you have. The last email in the thread, your notes, the numbers. This is the line most people skip and the one that changes the answer most.
What good looks like
An example you’d be happy with, or the one you’d reject and why. One example beats three paragraphs of description.
What the brief looks like in practice.
A shape of request we see from owners most weeks. The picture shows what changes in the reply, and the two cards under it show the prompts that produced each one. Nothing clever in the second version. It just contains the information the first one left in the owner’s head.
Write a follow up email to a customer who hasn’t gotten back to me about a roofing quote.
Missing: who they are, how long it’s been, what you’ve already said, what you want them to do, and what you won’t offer to get it.
Write a follow up email to a homeowner. I quoted a full roof replacement 12 days ago and she said she’d talk to her husband. We’ve emailed twice since the site visit. Here’s the thread: [PASTE] Goal: get a yes or a no this week, not another “we’ll think about it.” Do not offer a discount or mention the two earlier emails. Under 110 words. Friendly, not pushy. Give me two versions. One asks a direct question she can answer in one word. One offers to come back out and walk the roof with both of them.
Which tool for which job?
They overlap more every month, and for an everyday brief any of them is fine. The differences that still matter are where each one lives and what it’s connected to. Checked against each company’s own pages on October 16, 2026; we re-check this table every quarter.
| Tool | Reach for it when | Where it lives | Watch for |
|---|---|---|---|
| ChatGPT | Everyday drafting, brainstorming, images, and apps from other companies that run inside the chat. | Its own app and site. | Confident on facts it doesn’t have. Ask it to say what it’s unsure of. |
| Claude | Long documents, careful writing, working through a problem step by step, and code. | Its own app and site, and inside some work tools. | Tends to be thorough. Set a length if you want it short. |
| Gemini | Work that already sits in Gmail, Docs, Sheets and Drive, and anything tied to Google search. | Inside Google Workspace and Android, plus its own app. | Best when your material is already in Google. The value is the connection, not the model. |
| Perplexity | Research questions where you need to see the sources, not just an answer. It searches the web and footnotes each answer. | Its own app and site. | Read the sources it cites. A citation is not a verification. |
| Copilot | Work that already sits in Outlook, Word, Excel, PowerPoint and Teams. | Inside Microsoft 365 and Windows. | Same rule as Gemini. The value is the connection, not the model. |
Checked against: Microsoft Learn, Microsoft 365 Copilot service description, updated August 5, 2026 · Google Docs Editors Help, Gemini in Google Sheets · Perplexity, How does Perplexity work? · TechCrunch, October 24, 2025, on apps inside ChatGPT. If something here is out of date, tell us.
The first answer is a draft.
People who get good results from AI don’t write better first prompts. They treat the first reply as the start of a conversation, and three moves cover most of it.
Point at the line
“The second paragraph is too formal. Rewrite just that one.” Quote the part. Don’t start over.
Ask what it assumed
“What did you assume about the customer that I didn’t tell you?” The gaps in your brief show up here.
Make it argue the other side
“Now tell me why this email won’t get a reply.” The weak spots it finds are usually real.
Where this stops working.
It won’t make a model know your business. A brief gives it your material for one task, and it forgets by the next one. When the same context has to be in every answer, every day, across a team, that’s no longer a prompt problem. It’s a system, and that’s a different conversation.
It won’t catch a wrong fact for you. Reading the output is the job. A model will restate a wrong number in a confident sentence, and the sentence will read fine.
And it won’t tell you whether a task should be automated at all. That’s a different question, and we built the Bottleneck Finder™ for it, a seven-question diagnostic that sometimes concludes “build nothing.”
Questions people ask about prompts
What is a prompt, really?
A prompt is a brief. What you want, who it's for, what it must not do, the shape of the result, the material it should use, and what good looks like. The wording matters far less than whether those six things are in it.
Should I hand this task to AI at all?
Only if you can say in one sentence what a good result looks like. If you can't describe it, you can't check it, and an unchecked answer is worse than no answer. Work the problem by hand once, then come back.
Why does AI get my business wrong?
Because it only knows what you paste in. Facts about your customers, your numbers and your past work live with you, and if you don't hand them over the model invents the gap. Give it your material before you ask.
Which AI tool should I use?
For an everyday brief, any of the major ones. The differences that still matter are where each tool lives and what it's connected to: Gemini inside Google Workspace, Copilot inside Microsoft 365, Perplexity when you need to see the sources, ChatGPT and Claude for general drafting and longer work.
Does this page score my prompt?
No. A number on a prompt with no result to compare it against can't be defended, and the skill isn't phrasing anyway. The skill is writing a brief, and a brief is something you can check yourself with the six lines above.
Is the same brief getting typed ten times a week?
Then someone on your team is doing by hand what a system should hand them. Three minutes and seven questions, and you’ll see where the hours are going and what fixing it would run. If the honest answer is “do nothing,” it says so. No worries either way.
no email, no signup