Generative AI models (such as ChatGPT) are extremely powerful, but they only produce good results when they receive good instructions. In other words: the quality of the output depends directly on the quality of the prompt.
This document sets out a set of “Prompt Best Practices” to help anyone design clear, repeatable, and efficient instructions, both for simple tasks and for complex workflows.
1. What is really a “good prompt”?
A good prompt is much more than a question. It is a structured instruction that:
Defines a clear objective (what you want to achieve).
Provides enough context (what the AI does and does not know).
Specifies the AI’s role (how it should behave).
Describes the output format (how you want the answer to be delivered).
Includes constraints and quality criteria (what it should and should not do).
Is easy to reuse and adapt.
Think of the prompt as a briefing for a professional: the better that briefing is written, the better the work you will get.
2. Key principles of best practices
2.1. Objective first, always
Before writing a single word of your prompt, ask yourself:
What exactly do I want to achieve with this interaction?
What is the final output I need (a report, a list, an idea, an email, a project plan, etc.)?
Example (vague):
“Tell me about customer loyalty.”
Example (clear):
“Generate a list of 10 concrete ideas to improve customer loyalty in a B2C contact center, focused on reducing churn by 10% within 6 months.”
The more concrete the objective, the more actionable the response.
2.2. Provide relevant context
The AI does not know your organization, your product, or your business reality unless you explain it in the prompt.
Include context such as:
Type of company / industry.
Type of customer / end user.
Knowledge level of the target audience.
Channels involved (phone, email, chat, etc.).
Constraints (legal, regulatory, time, budget).
Example:
“We are a health insurance company that sells mainly by phone to end customers between 30 and 55 years old in Spain. The sales team has medium experience (2–3 years). I need…”
The more “situated” the AI is, the more relevant and realistic its recommendations will be.
2.3. Define the AI’s role
Asking “Explain this to me” is much less powerful than asking “Act as…”.
Examples of roles:
“Act as a senior CX consultant.”
“Act as a customer service trainer specialized in contact centers.”
“Act as a business analyst focused on data and reporting.”
“Act as a professional B2B copywriter.”
This sets expectations and raises the level of the response.
2.4. Specify the output format
Explicitly indicate how you want the result:
Numbered list.
Table with specific columns.
Step-by-step sequence.
Paragraphs of approximate length.
Checklist bullets.
Pseudocode or document structure.
Examples:
“Respond in table format with these columns: Step, Description, Owner, Estimated Time.”
“Limit your answer to 2 paragraphs and a list of 5 bullets.”
“Return the result as a training script with sections and subsections.”
This makes the content directly usable without having to reformat it afterwards.
2.5. Include quality criteria and constraints
Do not assume the AI understands your standards. State them clearly:
Level of detail: high, medium, very concise.
Tone: formal, friendly, technical, explanatory.
Language: neutral Spanish, Spanish (Spain), English, etc.
Scope: “Do not use examples from the United States, only Europe/ES.”
What to avoid: excessive jargon, legal terms, etc.
Example:
“Use a formal but approachable tone, aimed at managers. Avoid unnecessary technical jargon and do not use examples from outside Spain.”
2.6. Provide examples (and counterexamples)
When dealing with communication styles, scripts, or specific formats, examples are extremely valuable:
Example of what you want (positive).
Example of what you do NOT want (negative).
Example:
“I want an email similar to this style (example…). I do not want anything like this other one (example…), which is too aggressive.”
This significantly reduces iteration and rework.
2.7. Think in iterative mode
The first prompt will rarely be perfect. The normal way of working is in cycles:
Well-structured initial prompt.
Evaluation of the result.
Prompt adjustment (more context, more constraints, examples).
Refinement of the output.
Seeing the interaction as a continuous improvement dialogue is key: the AI does not get offended if you ask it to redo, correct, or restructure.
3. Recommended prompt template (“Prompt Best Practices”)
You can use this base template and adapt it case by case:
Role: “Act as [expert role].”
Objective: “Your goal is [specific result I need].”
Context: “Context: [brief description of company, audience, use case, relevant constraints].”
Task: “Task: [what you must do exactly: analyze, summarize, propose, write, correct, etc.].”
Output format: “Deliver the answer in [type of format: table, list, numbered sections, etc.].”
Tone and style: “Use a [formal/informal], [technical/explanatory] tone aimed at [audience profile].”
Constraints: “Take into account: [length limits, things to avoid, language, regulations, etc.].”
Examples (optional): “Example of what I am looking for: [text].”
Complete example:
“Act as a senior customer experience consultant in contact centers.
Your goal is to design a welcome call script for new customers of a health insurance company in Spain.
Context: we sell health insurance policies to customers aged 30 to 55, mainly via phone. We want to reduce complaint calls during the first month. Agents have 1 to 2 years of experience.
Task: create a detailed script for the first welcome call, including greeting, needs assessment, brief explanation of key coverages, and closing with an offer of additional support.
Output format: structure the script into sections with headings and bullets.
Tone and style: professional, empathetic, clear, and free of medical jargon.
Constraints: maximum 800 words. Do not promise anything that is not explicitly stated in the policy. The script must be easy for different agents to adapt.”
4. Prompt examples: weak vs improved
Example 1: Writing
Weak prompt:
“Write an email for a client.”
Problems:
No client context.
No clear objective.
No tone or format.
Improved prompt:
“Act as a B2B copywriter.
Write a short email (max. 150 words) for a client who has been with us for 6 months using a technical support service.
Objective: thank them for their trust and propose a brief review meeting to identify improvements.
Context: we are a ticketing software company for contact centers. The recipient is a Head of Operations of a contact center with 80 agents.
Tone: professional, direct, and strongly value-oriented.
Format: suggested subject line + email body in Spanish as spoken in Spain.”
Example 2: Analysis and synthesis
Weak prompt:
“Summarize this text.”
Improved prompt:
“Act as a business analyst.
Summarize the following text in a maximum of 10 clear, actionable bullets focused on recommendations to improve customer experience in a contact center.
Do not repeat specific examples; keep only the principles and best practices.
The summary must be in Spanish, in a professional tone, and aimed at managers.”
Example 3: Training and coaching
Weak prompt:
“Give me training ideas for agents.”
Improved prompt:
“Act as a customer service training specialist.
Create a 4-module training program for contact center agents who handle phone and chat interactions.
Objective: improve skills in active listening, objection handling, and closing conversations.
Context: agents have medium experience (1 year) and serve B2C customers of a telecommunications company in Spain.
Format: for each module include the title, learning objectives, estimated duration, and practical activities.”
5. Common mistakes when designing prompts
Being too generic
“Help me with this” is not enough. Always define objective and context.Not specifying the output format
This forces you to do a lot of editing afterwards.Forgetting to define the AI’s role
If you do not indicate the role, the model will tend to give answers that are too generic.Mixing several tasks in the same prompt
Example: “Analyze this text, rewrite it, suggest improvements, and create a presentation.” Better to split into steps.Asking for impossible or non-verifiable things
Example: “Guarantee that this will reduce churn by 30%.” The AI can suggest, not guarantee.Not reviewing critical outputs
Never use AI responses without human validation in areas with legal, medical, financial, or regulatory impact.Sharing sensitive data unnecessarily
Avoid including personally identifiable information (full names, IDs, phone numbers, etc.) when it is not essential to the task.
6. Quick “Prompt Best Practices” checklist
Before sending your prompt, check:
Have I clearly defined the objective?
Have I provided enough context (industry, audience, situation)?
Have I assigned an appropriate role to the AI?
Have I precisely described the task it must perform?
Have I specified the output format I need?
Have I indicated the tone, language, and level of detail?
Have I added relevant constraints and quality criteria?
Could I improve the prompt with one or two examples?
Is there anything sensitive (personal data, confidential information) I should remove or anonymize?
Am I prepared to iterate and adjust the prompt after seeing the first result?
If most of the answers are “yes”, your prompt is in the best-practice zone.