How to write prompts that actually work
6 Min. LesezeitA model doesn't read minds and doesn't guess at context. It does what the text says — and if the text is vague, so is the answer. Here's what a prompt that lands on the first try is made of.
Role: who the model should be
- “You are the art director of a branding agency” works better than “design a logo”. The role sets the vocabulary, the level of detail, and what the model even thinks to mention.
- Without a role the model answers like anyone would — averaged out. With one it reaches for specifics: an art director talks about grid and type size, a marketer about funnel and offer.
- Sharpen the role with experience: “a designer with ten years in packaging” gives a different answer than “a designer”.
Task: what exactly to do
- Bad: “write some copy for the site.” Good: “write a hero headline of no more than 60 characters.”
- Use one unambiguous verb: break down, come up with, rewrite, compare. “Help me with the copy” isn't a task, it's an invitation to chat.
- If there are several tasks, split and number them. The model will work through them in order and lose nothing.
Context: what the model can't know without you
- Product, audience, platform, constraints. Everything you hold in your head and treat as obvious does not exist for the model.
- The most common mistake is giving no context and then being surprised by a generic answer. The model won't ask — it will fill in the blank.
- Context is easiest to give as a labelled list: “Product: {what}. Audience: {who}. Platform: {where}.” That way it's easy to change without rewriting the prompt.
Format: what the answer should look like
- “Give me a list of five, each with a heading and one sentence of explanation.” Without this the model decides for you, and decides badly.
- Format also caps length. “No longer than 500 characters” saves you a minute of reading on every answer.
- It helps to say what shouldn't be there: “no preamble”, “don't explain what you're doing”, “straight to it”.
What kills a prompt
- Politeness. “Could you please possibly” doesn't improve the answer, it blurs the task.
- Several different tasks in one paragraph. The model does the first and smears the rest.
- Judgement words with no criterion: “nice”, “professional”, “catchy”. Replace them with something checkable: “under 60 characters”, “no corporate speak”, “a number in the first line”.
- No example. If you know what a good answer looks like, show one. A single example replaces three paragraphs of explanation.
A check before you send
- Read the prompt as someone who doesn't know your project. Is everything clear? If not, the model won't get it either.
- Is there a single word in it that could be read two ways? Remove it or pin it down.
- Can you tell from the prompt what the finished answer will look like? If not, add the format.