AI prompt best practices

These are the practices that consistently produce better output across ChatGPT, Claude, and Gemini. Not theory — actual techniques that change results.

2. 01

   Assign a role before the task

   Start every prompt with who the AI should be. "You are a senior product manager with 10 years of B2B experience" sets vocabulary, detail level, and assumptions before the task even begins. Without a role, the model defaults to a generic assistant voice that suits nobody in particular.
3. 02

   Be explicit about output format

   State the format you want before asking the question. "Return a bulleted list of 5 items, each under 20 words" is not over-specifying — it is removing ambiguity that would otherwise produce an output you have to reformat. Specify: list vs. prose, headers vs. no headers, length, and structure.
4. 03

   Front-load the most important information

   AI models pay more attention to content at the beginning and end of a prompt than the middle. Put the task and the most important constraints first. Instructions buried in the middle of a long prompt are more likely to be missed or given less weight.
5. 04

   Use negative constraints liberally

   "Do not use bullet points", "Do not start with a question", "Do not hedge with phrases like it is worth noting" — these negative instructions are extremely effective. They short-circuit the model's most common filler patterns and force more deliberate choices.
6. 05

   Provide the context the model cannot infer

   Who is the audience? What decision will this output inform? What has already been tried? What tone is appropriate for this situation? Anything the AI cannot know without being told should be in the prompt. Every piece of context you omit gets replaced by a guess.
7. 06

   Separate instructions from content with clear markers

   When you are pasting content (a document, data, an email) into a prompt with instructions, separate them clearly. Use labels like "INSTRUCTIONS:", "CONTENT:", "MY DATA:", "EXAMPLE OUTPUT:". This prevents the model from treating your instructions as part of the content to process.
8. 07

   Iterate rather than restart

   The first output is a starting point. Follow up with specific refinements: "The tone is too formal — rewrite the second paragraph", "This is 200 words too long — cut without losing the main argument", "The third point is vague — be more specific." Each iteration teaches the model more about what you need.
9. 08

   Ask for alternatives, not just one answer

   For writing, subject lines, headlines, and any output where taste matters: ask for 3-5 variations. You will almost always prefer one, or find the best elements split across two. "Give me 5 versions of this headline" costs nothing and dramatically improves your chances of getting something you can use.