Prompts that work

The model's earlier replies are sent back in each turn. That is how it "remembers" the chat. Your custom instructions, or project instructions, are your own part of the system prompt. With an API, you write all three roles yourself (Using AI through an API).

The model reads every role as one long stream of tokens. Models are trained to give system rules extra weight. But roles are labels on text, not walls. Users can often get a model to show its system prompt.

Keep your instructions, your examples and your material clearly apart. Label each part. Or wrap each one in simple tags, such as <instructions>, <example> and <feedback>. With long documents, a common tip is to put the documents first and your question at the end.

Tags help the model tell data from instructions. But text inside a document can still change what the model does. No wording makes it safe. Real limits belong outside the model. Learn more in Prompt injection.

• 🎭 Give it a role. "You are an experienced editor" changes its words, tone and focus. A role does not add knowledge the model lacks.
• 📋 Ask for a format. Ask for bullets, a table or a fixed template. Code can read structured output, such as JSON. A clear format is also easier to check.
• 🪜 Break it down. One prompt that does four jobs does each job worse. Split a big job into steps. Feed each result into the next. This is called prompt chaining.
• 🧠 Let it think first. Ask it to work through the problem before it answers. For example: list the themes, then count them. Reasoning models do this on their own. See Choosing a model.

A model is built to give a smooth answer. If your prompt suggests there must be an answer, it may make one up. See When AI makes things up.

Three short lines help a lot:

• "Use only the material provided."
• "If the answer is not there, say you don't know."
• "Quote the sentence you used."

Asking it to list what it assumed helps too. It can still be wrong. But now you can see where it is unsure, and check.

Priya's better brief said who the summary was for. It asked for themes, counts and one real quote each. It said to use only these comments. The new answer showed that export problems were a quarter of all complaints.

A better prompt is the cheapest fix. It costs nothing and works with any model. Fix the brief before you pay for more. See Prompt, RAG or fine-tune?

To practice more, try Anthropic's free interactive prompt engineering tutorial. Or read the community Prompt Engineering Guide. Both teach skills that work with other models too.