Claude is trained to be a helpful, honest, and harmless assistant. It is used to speaking in dialogue and you can instruct it in regular English.

The quality of the instructions you give Claude can have a large effect on the quality of its outputs, especially for complex tasks. This guide to prompt design will help you learn how to craft prompts that produce accurate and consistent results.

Claude works by sequence prediction

Claude is a conversational assistant, based on a large language model. The model uses all the text that you have sent it (your prompt) and all the text it has generated so far to predict the next token that would be most helpful.

This means that Claude constructs its responses one set of characters at a time, in order. It cannot go back and edit its response after it has written it unless you give it a chance to do so in a subsequent prompt.

Claude can also only see (and make predictions on) what is in its context window. It can't remember previous conversations unless you put them in the prompt, and it can't open links.

What is a prompt?

The text that you give Claude is designed to elicit, or "prompt", a relevant output. A prompt is usually in the form of a question or instructions. For example:

Human: Why is the sky blue?

Assistant:

The text that Claude responds with is called a "output".

The sky appears blue to us due to how the molecules in the air scatter light from the sun. The shorter wavelength light (blue) scatters more easily than the longer wavelength light (like reds and yellows), so when we look up at the sky, what we see is all the scattered blue light. This is known as Rayleigh scattering. The color of the sky varies based on conditions like the angle of the sun and amount of dust/water in the air, but during the day with a clear sky, the blue wavelength of light is most responsible for the color we perceive.

Human: / Assistant: formatting

Claude is trained to fill in text for the Assistant role as part of an ongoing dialogue between a human user (Human:) and an AI assistant (Assistant:).

Prompts sent via the API must contain \n\nHuman: and \n\nAssistant: as the signals of who's speaking. In Slack and our web interface we automatically add these for you.

For example, in the Console or in Claude-in-Slack, you can just ask Claude:

In one sentence, what is good about the color blue?

And it will respond:

Blue is often seen as a calming and soothing color.

If you send the same prompt to the API, it may behave in unexpected ways, like making up answers well beyond what was asked for in the prompt. This is because Claude is trained to fill in text for the Assistant role as part of an ongoing dialogue between a human user (Human:) and an AI assistant (Assistant:). Without this structure, Claude doesn't know what to do or when to stop, so it just keeps on going with the arc that's already present.

The prompt sent to the API should be:

Human: In one sentence, what is good about the color blue?

Assistant:

💡

Why?

Claude has been trained and fine-tuned using RLHF (reinforcement learning with human feedback) methods on \n\nHuman: and \n\nAssistant: data like this, so you will need to use these prompts in the API in order to stay “on-distribution” and get the expected results. It's important to remember to have the two newlines before both Human and Assistant, as that's what it was trained on.

Prompt length

The maximum prompt length that Claude can see is its context window. Claude's context window is currently ~75,000 words / ~100,000 tokens / ~340,000 Unicode characters.

Right now when this context window is exceeded in the API Claude is likely to return an incoherent response. We apologize for this “sharp edge”.