Ten prompting habits that work, from the official guides

You do not need secret formulas to get good work out of an AI model. Anthropic and OpenAI both publish guidance on how to brief their models, and the advice overlaps more than you might expect. Anthropic’s guide starts with an image worth keeping: treat the model as a brilliant but new employee who lacks context on your norms and workflows. Almost every habit below follows from that.

One thing first. Anthropic’s guide assumes you know what a good result looks like and have a way to check it. Before you polish a prompt, write down in a sentence or two what success means.

1. Brief it like a capable new colleague

Anthropic’s golden rule: show your prompt to a colleague with minimal context on the task. If they would be confused, the model will be too. One sentence about the role it should play also helps. Instead of “Write a newsletter intro”, try “You write for our B2B newsletter: write a 60-word intro for existing clients announcing that our support desk now opens at 8:00.”

2. Say why, not only what

Explaining the reason behind an instruction helps the model understand your goal, and it generalises from the explanation. Anthropic’s own example replaces “NEVER use ellipses” with the reason: the text will be read aloud by a text-to-speech engine that cannot pronounce them. Instead of “Keep it short”, try “Keep it short, because it appears as a notification on a phone lock screen.”

3. Swap vague words for numbers and formats

OpenAI advises being specific about context, outcome, length, format and style, and cutting “fluffy”, imprecise descriptions. Instead of “A fairly short product description, not too long”, try “A product description of 3 to 5 sentences, addressing the reader as you, ending with the price.”

4. Ask for the result, not for suggestions

Models take your wording literally. Anthropic notes that “Change this function to improve its performance” works better than “Can you suggest some changes”, and that if you want more than the basics, you have to ask for it. Instead of “Any ideas for this headline?”, try “Rewrite this headline in three versions: one factual, one playful, one built around a number.”

5. Say what to do, not only what to avoid

Both guides make this point: a list of prohibitions leaves the model guessing what you do want. Instead of “Don’t use jargon”, try “Write so that a new customer without a technical background understands every sentence.”

6. Show three to five examples

A few well-chosen examples improve accuracy and consistency. Anthropic recommends three to five, close to your real use case and varied enough that the model does not copy an accidental pattern; OpenAI also asks for a diverse range of inputs. Instead of “Write subject lines in our style”, try “Here are four subject lines that worked well for us; write five new ones in the same style for the spring sale.”

7. Keep instructions and material apart

When instructions, background and source text run together, the model can mistake one for another. Anthropic suggests wrapping each part in its own XML-style tag with a clear name; OpenAI suggests separators such as triple quotes, or Markdown headings. Instead of pasting a report straight under your question, try “Summarise the report below in five bullets”, followed by the report inside clearly named tags.

8. Long material first, question last

For long documents, Anthropic advises putting the material at the top and your question at the end; in its tests, this improved response quality by up to 30 percent, especially with several documents. Asking the model to quote the relevant passages first helps it focus. Instead of “What does this contract say about termination?” above forty pages, put the pages first and end with “First quote the passages about termination, then explain in plain words what they mean for us.”

9. Allow “I don’t know”, and ask for a check

Anthropic’s guidance on reducing hallucinations starts with explicitly allowing the model to say it does not know. Ask it to cite a source for each claim, or to verify its answer against your criteria before it finishes. Instead of “What is the market share of our three competitors?”, try “Use only the reports I attached, cite the source of every figure, and say so when a figure is not in them.”

10. Give thinking models a goal, not a script

Many current models reason before they answer. OpenAI compares a reasoning model to a senior colleague you can give a goal, and a classic GPT model to a junior colleague who needs explicit instructions. Anthropic adds that a prompt like “think thoroughly” often produces better reasoning than a hand-written step-by-step plan. Instead of a ten-step recipe for an analysis, try “Find out why our March conversions dropped; check the data first, then tell me what you would do.”

Test, then share what works

Try a prompt on several real cases before you trust it, and keep the version that wins. OpenAI also recommends building evaluations, so you notice when a new model version behaves differently.

A prompt that works is company knowledge. Treat it that way.

In Zaela, a method that works is written down once as a skill and used by everyone, next to your brand rules and tone of voice in one shared memory. Ask for the same thing twice? Turn it into an agent.