🔴CohereBeginner

Cohere Prompt Engineering Guide

Best practices for prompting Cohere's Command models effectively

View Official Documentation

TL;DR

•Be clear, concise, and specific
•Include examples of desired output format
•Use delimiters with explanatory headers
•Place instructions at the beginning
•Start the completion yourself to guide format
•Split complex tasks into smaller prompts

Key Principles

1Clarity and specificity are paramount

2Delimit different information types with headers

3Provide context and background when helpful

4Use examples for complex formatting

5Tell the model what to do, not what to avoid

6Control length explicitly when needed

Clear and Specific Prompts

The most effective prompts are clear, concise, specific, and include examples of exactly what a response should look like. Vague prompts lead to unpredictable outputs.

Examples

Be specific about requirements

Bad:

Write a product description

Good:

Write a product description for wireless earbuds.
- Length: 50-75 words
- Tone: Enthusiastic but professional
- Include: Key features, target audience, call-to-action
- Avoid: Technical jargon

Specific requirements ensure the output matches marketing needs.

Use Delimiters and Headers

Place instructions at the beginning of the prompt. Delimit different types of information (instructions, context, resources) with explanatory headers to help the model understand the prompt structure.

Examples

Structure with clear headers

Good:

## Instructions
Summarize the following article in 3 bullet points.

## Article
{article_text}

## Output Format
- Bullet point 1
- Bullet point 2
- Bullet point 3

Headers clearly separate instructions from content and expected output.

#prompt-engineering#cohere#command#few-shot#delimiters#task-splitting

Last updated: December 1, 2024

Provide Examples

For complex formatting or specific output styles, showing examples is more effective than describing them. Let the model learn the pattern from your examples.

Examples

Show don't tell

Good:

Convert customer feedback to structured data.

Example:
Feedback: "Love the app but it crashes sometimes on my iPhone"
Output: {"sentiment": "mixed", "topic": "stability", "device": "iPhone", "action": "investigate crashes"}

Now convert:
Feedback: "The new update is amazing, especially the dark mode!"

The example shows exact JSON structure and field extraction logic.

Do vs Don't Instructions

Tell the model what TO DO rather than what NOT to do. Positive instructions are clearer and more reliably followed than negative ones.

Examples

Positive instructions work better

Bad:

Don't use complex words. Don't write more than 100 words. Don't be formal.

Good:

Use simple, everyday language. Keep your response under 100 words. Write in a casual, friendly tone.

Positive instructions clearly define the target behavior.

Begin the Completion

You can guide the model's output format by starting the completion yourself. This is especially useful for ensuring specific formats like JSON or lists.

Examples

Start the response format

Good:

List 3 benefits of exercise.

Benefits:
1.

Starting with "1." ensures a numbered list format.

Task Splitting

For complex tasks, split them into smaller, sequential prompts. Use the output of one prompt as input to the next. This improves accuracy and makes the process more controllable.

Examples

Break down analysis tasks

Good:

Task 1: Extract all company names mentioned in this article.
Task 2: For each company, identify their industry.
Task 3: Summarize the relationship between the companies.

Let's start with Task 1.

Sequential tasks ensure each step is completed accurately before proceeding.