Recette parfaite de prompt ChatGPT en 10 composants

A perfect ChatGPT prompt has 10 components: Objective (15%), Role (10%), Context (20%), Input Data (15%), Quality Checks (4%), Constraints (8%), Examples (5%), Iteration Request (5%), Instructions (10%), and Output Format (8%). You do not need all 10 every time, but knowing which levers to pull changes the game. 1. Context (20% of the impact): This is the background reality the model needs to inhabit. Include your business type, industry, specific audience, goals, and current challenges. Never assume the model knows your situation. Pro tip: Write your context once, save it in a text file (or as Custom Instructions/Project knowledge), and paste it in every time. Example: "My company sells project management software to remote teams with 10 to 100 employees. Our main challenge is that buyers think we are too expensive compared to free tools." 2. Objective (15% of the impact): This is the clear definition of the task. AI performs best when goals are explicit, measurable, and bounded. Pro tip: Replace vague verbs with specific outcomes. Do not say "help me with." Say "create," "diagnose," or "rewrite." Bad: "Tell me about marketing." Good: "Create a 90-day content marketing strategy for a SaaS startup targeting small businesses." 3. Input Data (15% of the impact): Hand over the actual information the model needs to do the work. This could be meeting notes, customer feedback, a rough draft, a research report, or website copy. Pro tip: Use XML tags (like and ) to separate your input data from your instructions. Example: "Here are the raw transcripts from three customer interviews. Based on these transcripts..." 4. Role (10% of the impact): Tell the model who it should be. Assigning a role activates completely different knowledge clusters and reasoning patterns. Pro tip: Pair the role with a specific tone or philosophy to narrow the focus. Example: "Act as a world-class direct response copywriter who specializes in concise, punchy, David Ogilvy-style email campaigns." 5. Instructions (10% of the impact): Tell the AI exactly what to do with the Context, Objective, and Input Data. Use strong action verbs. Pro tip: Break complex instructions into numbered steps. Example: "1. Analyze the data. 2. Identify the three most common complaints. 3. Prioritize recommendations to fix them. 4. Explain your reasoning." 6. Constraints (8% of the impact): Set the boundaries. Include maximum word counts, reading levels, budget limits, or things it is not allowed to do. Pro tip: Negative constraints are powerful for killing the "AI smell." Example: "Maximum 500 words. Do not use the words 'delve,' 'crucial,' or 'tapestry.' Keep the reading level at an 8th-grade standard. Use only the provided information." 7. Output Format (8% of the impact): Specify exactly what shape the answer should take. Pro tip: If moving data into another system, ask for CSV or JSON. If presenting, ask for a Markdown table. Example: "Present the answer in a table with three columns: Problem, Impact, and Proposed Solution." 8. Examples (5% of the impact): Show the model what good output looks like. Providing an example of the input, the desired output, and the format reduces misinterpretation. Pro tip: If the model keeps failing, giving one perfect example is usually faster than rewriting instructions. Example: "Here is an example of the tone I want. Input: Customer complains about pricing. Output: Highlight ROI and provide three relevant case studies." 9. Iteration Request (5% of the impact): Prompting is a back-and-forth conversation. Build iteration into the prompt. Pro tip: Ask for multiple options to choose the best direction. Example: "Generate three distinct alternatives for the headline. Then, critique your own responses and tell me which one is strongest and why." 10. Quality Checks (4% of the impact): Ask the AI to verify its own work before finalizing. Pro tip: Add a quality check to the end of any complex analytical prompt. Example: "Before finalizing your answer, check for factual accuracy, identify any weak assumptions you made, and highlight any missing information that would make your recommendation stronger." Remember: The prompt you type is a container. If you only fill the Instructions section, the model has to guess the rest. Fill the container, and the model stops guessing and starts working.

Recette parfaite de prompt ChatGPT en 10 composants

Contenu du PromptEnregistrer

Utilisation

Références

Discussion