How to ask AI for things in 2026

What you'll see in this lesson

1. The three official guides say the same thing

   OpenAI, Anthropic and Google publish guides for asking their current models for things, and they agree on the essentials.

2. Start with the result and how you'll know it's good

   A good request names the deliverable, the audience and the criteria you'll use to judge the result.

3. Explain the why: it beats a rule in capital letters

   If the model knows what each instruction is for, it applies it better, even in cases you hadn't foreseen.

4. Limits and format: say what you want, not what you don't

   Set the real limits and describe the output format in positive terms: it works better than a list of bans.

5. A good example teaches more than ten instructions

   Showing a few examples of the result you want is one of the most reliable ways to set tone, format and structure.

6. XML tags: separate instructions, data and examples

   Wrapping each part in a tag like <document> or <task> keeps the model from confusing your data with your orders.

7. 2023 tricks that are useless today (or get in the way)

   Threats, tips, "you're the world's best expert" and rules in capital letters no longer improve answers.

8. When a very long prompt gives worse answers

   In OpenAI's internal tests, trimming the system prompt raised the score and used between 41 and 66% fewer tokens.

9. Summary: ask like you would a smart colleague

   Asking well in 2026 means describing the job as you'd explain it to someone capable who doesn't know your case.