By James Finley. Published Oct 13, 2026.
A better prompt is rarely a longer one. It's usually the right technique for the job. These six habits cover most everyday work in ChatGPT and Claude. Each one comes with a prompt you can copy and change, and each is checked against OpenAI's and Anthropic's own prompting guides.
Tell the AI who it is and who it's writing for. Anthropic's guide says that even a single sentence of role makes a difference.
Try this: "You are a customer success manager at a software company. Write a 3-sentence email to a frustrated client about a delayed feature. Apologize once, give the new date, and don't over-apologize."
A role sets tone and vocabulary. It doesn't fix a vague task, so still say what you want, how long it should be, and who will read it.
If every weekly report has to look the same, show the AI what "the same" looks like. OpenAI's guide says to show the format you want through examples. Anthropic suggests 3 to 5 examples, and in Claude you can wrap each one in <example> tags so it knows where each starts and ends.
Try this: "Follow this format exactly. Example: Shipped: the new intake form. Blocked: waiting on legal review. Next: training for the front desk. Now write this week's report from my notes below."
A simple rule: try the prompt without examples first. If the format comes back wrong, add a few examples that match your instructions closely. OpenAI notes that its newer reasoning models often don't need examples, and that examples which don't fit the instructions can make results worse.
One giant prompt that asks for research, a summary, a table and an email at once tends to do each part poorly. Ask for one part, check it, then ask for the next.
You may have heard the tip to add "think step by step." OpenAI now says this is unnecessary for its reasoning models, because they already work through the problem before answering, and it can sometimes make results worse. What still helps is asking the AI to show its working so you can check it: "Give me the formula to count orders over $500 in column B, and say which cells it uses."
For a task you repeat, describe the result you want and let the AI draft the prompt. This is often called meta-prompting.
Try this: "I need a prompt I can reuse every month to summarize our employee survey for managers who aren't technical. It should ask for a 200-word summary, three themes and one suggested action. Write the prompt for me."
Both OpenAI's and Anthropic's developer guides point to tools that draft prompts this way. Read the prompt it gives you, fix anything that's off, and save it. Skip this for quick one-off jobs like a subject line, where writing the prompt yourself is faster.
OpenAI's help page for ChatGPT Images says to be specific about the details that matter, and to include any words the image should show. Cover four things:
Try this: "A flat-design poster for a charity 5K run. Runners crossing a finish line in the lower half, open sky above. Teal and orange. Include the text 'Charity 5K, April 12' in a clean sans-serif font."
Once the image is made, you can ask for changes in the same chat, such as "make the headline larger," instead of starting over.
Most advice works in both tools. Two points differ:
<instructions> and <examples> for longer prompts. In ChatGPT, clear headings or the ### separators do the same job.Both guides agree on the basics: be specific, give exact targets instead of words like "fairly short," and say what to do, not only what to avoid.
AI tools can invent names, dates, numbers and policy details. Before anything goes out, ask: "List every claim in this draft that needs checking." Then check each one yourself, and cut anything you can't confirm.
Keep your company's rules in mind too. Only paste work information into an AI tool your employer has approved.
Reading about these techniques is a start. The habit forms when you write prompts, see where they fall short and fix them. Our Prompting Fundamentals course walks through the building blocks of a clear prompt, repairing a vague one, and making a reusable template, with hands-on exercises. Its preview lessons are free, with no account needed.