Notre avis

Ce guide fournit des principes et une structure pour créer des instructions système efficaces destinées aux agents et assistants IA.

Points forts

Principes clairs et actionnables (ex. : dire ce qu'il faut faire, pas ce qu'il ne faut pas faire).
Structure de prompt recommandée (rôle, responsabilités, contraintes, format).
Exemples concrets illustrant chaque principe.
Tableau des pièges courants avec solutions.

Limites

Ne couvre pas les techniques avancées comme le few-shot learning ou le chain-of-thought.
Les exemples sont génériques et peuvent nécessiter adaptation pour des cas très spécifiques.
Suppose que l'utilisateur a déjà une bonne compréhension du domaine de l'IA.

Quand l'utiliser

Lorsque vous devez rédiger ou améliorer un prompt système pour un agent IA afin d'obtenir des réponses cohérentes et de haute qualité.

Quand l'éviter

Si vous cherchez des techniques avancées de prompt engineering (ex. : few-shot, CoT) ou si le comportement souhaité est très simple et ne nécessite pas de structure formelle.

Exemples

Write a system prompt for a code review assistant

Help me create a system prompt for an AI assistant that reviews pull requests. It should focus on security, performance, and best practices, and output comments in a specific format.

Improve an existing system prompt

I have this system prompt: 'Be helpful and concise.' How can I improve it using the principles from prompt-creator? Rewrite it with explicit instructions and constraints.

Apply prompt structure guidelines

I need to write a system prompt for a customer support bot. Use the recommended structure (role, responsibilities, constraints, output format) and include an example.

name: prompt-creator description: Guide for creating effective system prompts for AI agents and assistants. Use when writing, reviewing, or improving system prompts, agent instructions, or assistant configurations.

Prompt Creator

Guidelines for designing effective system prompts for AI agents and assistants.

Core Principles

Start Simple, Iterate

Begin with minimal instructions and add complexity only as needed. Prompt design is iterative—test with real inputs and refine based on results.

For complex agent behaviors, break the task into simpler subtasks first. Build up gradually rather than front-loading complexity.

Use Clear Instructions

Structure prompts with explicit commands: "Write", "Classify", "Summarize", "Translate", "Extract", "Analyze".

Place instructions at the beginning. Use separators to distinguish sections:

### Instructions ###
[Core behavior and responsibilities]

### Context ###
[Domain knowledge and constraints]

### Output Format ###
[Expected response structure]

Be Specific and Direct

Vague:

Keep responses short and helpful.

Specific:

Respond in 2-3 sentences. Include one actionable recommendation.

The more descriptive and detailed the prompt, the better the results. Include:

Exact output formats
Specific constraints
Concrete examples of desired behavior

Say What To Do, Not What Not To Do

Weak (focuses on prohibitions):

DO NOT ask for personal information.
DO NOT make assumptions.
DO NOT provide medical advice.

Strong (focuses on actions):

Recommend from the top 10 trending items.
Ask clarifying questions before proceeding.
Direct health questions to qualified professionals.

Positive instructions are more effective than negative constraints.

Provide Examples for Desired Formats

When a specific output structure is needed, show it:

Extract entities from the input text.

Desired format:
- People: <comma_separated_list>
- Places: <comma_separated_list>
- Organizations: <comma_separated_list>

Example input: "Dr. Sarah Chen presented findings at MIT yesterday."
Example output:
- People: Dr. Sarah Chen
- Places: MIT
- Organizations: MIT

Balance Detail with Relevance

Include only details that contribute to the task. Unnecessary context dilutes the signal.

Ask: "Does this information change how the agent should behave?" If not, remove it.

System Prompt Structure

Effective system prompts typically follow this order:

Role/Identity - Who the agent is
Core Responsibilities - Primary tasks and behaviors
Constraints - Boundaries and limitations
Output Format - How to structure responses
Examples - Concrete demonstrations (if needed)

You are a [role] that [primary function].

## Responsibilities
- [Responsibility 1]
- [Responsibility 2]

## Constraints
- [Constraint 1]
- [Constraint 2]

## Response Format
[Structure specification]

Common Pitfalls

| Pitfall | Problem | Solution | |---------|---------|----------| | Vague quantities | "a few", "short", "brief" | Use exact numbers: "2-3 sentences" | | Double negatives | "don't avoid" | Use positive framing | | Contradictory rules | "be concise" + "be thorough" | Prioritize or scope each rule | | Missing edge cases | Undefined behavior | Add fallback instructions | | Excessive length | Context dilution | Remove non-essential content |

Créateur de Prompts Efficaces

Recommandé pour