Prompts That Work with Any AI
From the First Question to the Professional Outcome
Alejandro Vargas
Free sample · First chapter
You wrote:
> “Give me ten ideas for improving my company’s communication.”
The AI responds with a correct, orderly and fairly useless list:
There is no obvious error. The response talks about communication, offers ten ideas and uses a reasonable tone. However, it probably was not what you were looking for. Perhaps you wanted specific actions for a company with eight employees, a small budget and professional clients. Or maybe you needed ideas for a launch campaign over the next thirty days.
The AI did not ignore your intention. It simply did not receive it.
This is the starting point for learning to write better prompts: an instruction does not contain what you are thinking while you write it. It contains the words you have put into it, along with the context the assistant can access. If an important part of your objective remains in your head, the tool has to fill in the gaps. And when it fills in gaps, it may choose an interpretation different from yours.
AI assistants generate responses from the texts they receive, the previous conversation and, depending on the tool, other documents or data you have provided. They do not automatically have access to your intentions, priorities or personal criteria.
When you write “make it more professional,” you know what you mean because you know the text and the situation. To you, “professional” might mean:
The AI does not know which of these options you prefer unless you tell it. It may choose a common interpretation, but a common interpretation is not necessarily the right one for your situation.
The same applies to expressions such as:
These are understandable instructions in a human conversation, but they are insufficient for obtaining a controlled result. They do not specify the objective, audience, limitations or form the response should take.
A vague request is not necessarily useless. It can help start a conversation or explore possibilities. The problem arises when you expect it to produce a specific final result directly.
Imagine asking an assistant:
> “Write an email to chase up an invoice.”
Several important details are missing:
The assistant may invent some details, leave them out or write a generic message that works for almost any situation. The last option may seem prudent, but it usually means you will have to rewrite the email.
The result is not poor because the AI is incapable of writing an email. It is poor because the instruction allows too many different responses.
You can think of a prompt as a professional brief. If you tell someone, “Prepare something about this matter,” they may ask you questions before starting. An AI assistant can also ask for clarification, but it will often try to answer immediately. That is why it is useful to include from the outset any information that would change the result.
EXAMPLE: FROM A VAGUE REQUEST TO A CLEAR BRIEF
Vague request:
> Write an email to chase up an invoice.
More precise brief:
> Write a reminder email addressed to a client company. Invoice 348, for €1,250, was due 15 days ago. This is the first reminder, so the tone should be cordial and professional, without threats. Ask them to confirm the expected payment date. Use a subject line and a body of no more than 150 words. Do not invent any details that are not included in the instruction.
The second version does not guarantee that the text will be perfect, but it greatly reduces the scope for misinterpretation.
Two people can ask for help on the same subject and receive very different responses. The difference lies not only in the topic, but also in the conditions of the assignment.
A complete instruction usually clarifies five aspects:
1. What you need to achieve.
Informing, persuading, comparing, summarizing and making a decision are not the same thing.
2. Who the result is for.
The language changes depending on whether the audience is a client, a management team, secondary-school students or people with no prior knowledge.
3. What information the AI should use.
This could be a text, a data set, a description of your business or the content of a document.
4. What limits it must respect.
Length, tone, language, budget, date, applicable legislation, available information or topics it should avoid.
5. How you want to receive the response.
Paragraphs, numbered steps, a table, a script, an email, a summary with sections or a prioritized list.
These elements are not decorative. Each one rules out possible interpretations.
For example, “compare these tools” leaves many questions open. Do you want to compare price, features, ease of use or security? For what type of user? Based on what available information? Do you need a final recommendation? In what format?
A better instruction might be:
> Compare these three tools for a four-person team that needs to draft documents and summarize meetings. Assess ease of use, collaboration, price and the ability to work in Spanish. Use only the data included in the text I will provide. If information is missing, say so. Present the comparison in a table and finish with a reasoned recommendation of no more than 100 words.
The first request opens an undefined investigation. The second defines a task.
Context is the information that makes it possible to interpret an instruction correctly. It may include details about your situation, the audience, the source material and the result you are seeking.
Without context, the assistant tends to rely on general responses. Not because it always wants to be superficial, but because it lacks the criteria needed to choose between several valid responses.
Notice the difference:
> Give me ideas for posting on LinkedIn.
Compared with:
> I run a small employment consultancy. Our audience consists of self-employed people and companies with fewer than 20 employees. I want to post for four weeks to attract enquiries, without using an aggressively sales-focused tone. Propose 12 ideas for LinkedIn. For each one, indicate the topic, format, problem it addresses and call to action. Avoid specific legal advice that may become outdated.
In the second instruction, the AI knows the industry, audience, channel, time frame, objective and limitations. It has more information to produce something useful and fewer reasons to fill the response with clichés.
You do not need to include every detail of your life. You need to include the details that could change the response. A simple way to decide is to ask yourself:
> If the AI does not know this detail, could it choose a solution I would reject?
If the answer is yes, include that detail in the prompt.
It is often recommended that prompts begin with formulas such as:
> Act as a marketing expert.
Assigning a role can guide the type of response, but it does not replace a specific instruction. The phrase by itself does not define the work to be done, the information the AI may use or the form the response should take.
Compare these two cases:
> Act as a human resources expert and write something about interviews.
And:
> Act as a careers adviser. Prepare a guide for people facing their first job interview. Explain how to answer five common questions, include an example answer for each one and avoid recommending that candidates invent experience. Use clear language and limit the guide to 800 words.
In the second case, the role provides context, but the result depends above all on the objective, audience and conditions.
It is also worth avoiding requests for an authority that could be misleading. “Act as a doctor” does not turn the AI into a healthcare professional or guarantee that the information will be appropriate for a clinical situation. A role helps focus the style and criteria of a task; it does not confirm knowledge, credentials or access to up-to-date data.
You can receive a correct response and still have to redo it because it is not organized in the way you need.
If you want to turn a meeting into tasks, it is not enough to ask:
> Summarize this meeting.
You may get several paragraphs with the main ideas. But perhaps you need an operational list with people responsible and deadlines.
A more useful instruction would be:
> Analyze the meeting transcript and extract only the agreed tasks. Present a table with these columns: task, person responsible, deadline and dependency. If a detail does not appear, write “not indicated.” Do not turn proposals that were not approved into tasks.
The format serves an important function: it forces the information to be separated and makes absences visible. If the “deadline” field appears as “not indicated,” you can review the meeting. In a narrative summary, that detail could remain hidden.
You can request formats such as:
The more specific the format, the easier it will be to review and reuse the result.
Many people believe that setting limits reduces AI creativity. In reality, a well-chosen constraint prevents the response from straying from your objective.
You can limit:
For example:
> Suggest five names for a course aimed at primary school teachers. They must be easy to remember, contain no anglicisms and have a maximum of six words. Avoid promises such as “learn effortlessly” or “master any subject.”
Constraints do not guarantee that you will like the names, but they make the proposals comparable. They also allow you to quickly identify when the AI has not followed the instruction.
There is a difference between a useful constraint and a confusing accumulation of instructions. If you include twenty contradictory conditions, the assistant may prioritize some and set others aside. It is better to write the requirements in order of importance and separate the main objective from secondary details.
IMPORTANT NOTICE: AI MAY FILL GAPS WITH FALSE INFORMATION
When data is missing, some assistants try to provide a fluent answer even when they do not have sufficient evidence. They may invent figures, sources, names, dates, or product features.
Do not avoid this risk with the phrase “be precise” alone. State what it should do when it cannot verify something:
> Use only the information provided. If a necessary piece of information is missing, say so explicitly and do not fill it in with an assumption.
Then review the important claims. A clear instruction reduces fabrications, but it does not replace human verification.
In a real task, there are often several objectives competing with one another. You want a text that is brief, complete, friendly, and forceful. You want many ideas, but you also want them to be easy to apply. You want a simple explanation without losing precision.
If you do not indicate what it should prioritize, AI will have to decide. You can help it by writing something like:
> Prioritize clarity over originality. If it is not possible to include all the details in 300 words, keep the practical instructions and remove the secondary examples.
Or:
> Give priority to low-cost options that can be implemented in less than a week. Then include the more complex alternatives.
This kind of instruction is especially useful when there are limits on space, time, or budget. It is not enough to list preferences: you have to indicate which one wins when they conflict.
A well-constructed prompt does not just improve the first response. It makes it easier to explain what went wrong.
If you write “make it better” after receiving an unsatisfactory result, AI still does not know which aspect it should change. Instead, you can point out the requirement that was not met:
> The content is suitable, but the text exceeds the 500-word limit. Keep the main ideas, remove repetitions, and retain the examples related to self-employed workers.
The conversation stops revolving around general opinions and begins to focus on observable criteria.
You can also ask for a review before the final response:
> Before drafting the text, identify the missing information and ask up to three specific questions. Do not start drafting until there is enough information.
This strategy is useful when an early wrong decision could force you to redo the entire task. For simple tasks, it may be faster to provide all the data from the outset. For important tasks, asking for clarification can save time.
Before writing a prompt, complete these sentences mentally:
- I need AI to…
Define the main action: draft, classify, explain, compare, review, or propose.
- The result will be used by…
Describe the intended person or group.
- It must take into account that…
Add the context that changes the solution.
- It must include…
List the essential elements.
- It must avoid…
Point out errors, approaches, or data that you do not want to appear.
- I want to receive it as…
Specify the output format.
- If information is missing…
Indicate whether it should ask questions, acknowledge the lack of data, or work with a clearly marked assumption.
Then organize the instruction. Start with the task, add the context, set the conditions, and finish with the output format. This is not a mandatory rule, but it usually makes the instruction easier to read.
A basic template could be:
> Task: [what it should do].
> Context: [information needed to understand the situation].
> Audience: [who will use or read the result].
> Requirements: [mandatory elements and limits].
> Format: [how it should present the response].
> Insufficient data: [what it should do if it cannot verify something].
You do not have to copy this structure literally every time. Its value lies in helping you check that you have not forgotten an important part of the assignment.
Review this prompt:
> Explain how I can organize my time better and give me a routine.
Before sending it, answer:
A more complete version could be:
> Help me organize my workday. I am self-employed and work from home, spending about six hours a day preparing estimates, answering emails, and attending video calls. Meetings usually take place in the morning, but they change from day to day. I want to reduce constant task switching and set aside at least one hour for focused work. Propose a flexible routine with time blocks, prioritization rules, and an alternative for days with many meetings. Explain it in no more than 600 words, and do not assume that I can control my clients’ schedules.
The second version does not describe an entire life. It provides only the information needed to design a response that more closely matches the real situation.
The central rule of this chapter is simple: AI cannot carry out an intention that you have not turned into observable instructions. The better you define the objective, context, limits, and format, the less AI will have to guess. And the less it has to guess, the easier it will be to obtain a useful response, review it, and repeat the process with another task or another assistant.
When you use an AI assistant, it is easy to write the first thing that comes to mind:
> “Give me ideas for promoting my business.”
The sentence seems clear because you know the business, the problem, and the result you expect. The AI, by contrast, receives only those eight words. It does not know what you sell, whom you are targeting, which channels you use, how much you can spend, or what you consider a good idea. That is why it may respond with proposals that appear sound but are too general to be useful.
The difference between a mediocre response and a useful one often lies in the quality of the brief. A precise brief turns a broad intention into a task with enough information to work from.
An improvised question often looks like this:
> “How can I improve my social media posts?”
A precise brief adds the information needed to guide the response:
> “Act as a content adviser for a small language academy. Propose ten Instagram post ideas aimed at adults who want to learn English for work-related reasons. Each idea must include a title, the objective of the post, and a call to action. Avoid generic motivational phrases and do not make up market data. Present the result in a table.”
The second version does not guarantee a perfect response, but it greatly reduces the number of decisions the AI has to make on its own. You have specified:
The AI cannot read the intention you do not write down. It can make assumptions based on common patterns, but those assumptions may not match your situation. The more important the result, the less advisable it is to leave these decisions open.
Context answers a simple question: what does the AI need to know to understand the brief?
You do not need to tell the whole story of your company or copy endless documents. Include the information that changes the response. For example, when requesting marketing copy, the following may be relevant:
Compare these two requests:
> “Write a description of my consulting service.”
> “Write a description of a tax advisory service for self-employed people who have just started their business. The text will be published on the home page of a website and must explain what problems it solves, how the initial consultation works, and what kind of assistance it provides. Use a clear, professional tone, without promising specific savings or guaranteed results. Maximum length: 150 words.”
In the first version, the AI has to invent almost the entire situation. In the second, it can focus on writing.
This does not mean you should always include every possible detail. The context should be sufficient and relevant. Adding information that does not affect the task can make the prompt longer without making it better.
Sometimes it is useful to indicate the perspective from which the AI should work:
> “Act as a mathematics teacher for secondary-school students.”
This kind of instruction can help adjust the vocabulary, depth, and approach. However, saying only “act as an expert” does not define the task. The role does not replace context or instructions.
A weak prompt would be:
> “Act as a marketing expert and tell me what to do.”
The assistant still does not know what kind of business you have, what goal you are pursuing, or what type of response you expect. A more complete prompt would be:
> “Act as a marketing consultant for a local shop selling organic products. Analyse these three promotional ideas and rank them according to their cost and ease of execution. Explain the reason for each position in a brief paragraph. Do not propose campaigns that require hiring additional staff.”
The role provides a perspective. The brief defines the job.
Many responses disappoint because people know what they want to read but do not describe it. “Make it professional,” “make it comprehensive,” or “give me something creative” are difficult instructions to measure. Different readers may interpret them in different ways.
It is better to define the result using observable details:
For example:
> “Summarise the document in five points. Each point must contain a concluding sentence and a second sentence with the data that supports it. Keep the original figures and label any information not included in the document as ‘not stated’.”
This instruction is more useful than “write a clear and comprehensive summary” because it establishes what the output should look like and what the AI should do when information is missing.
A simple way to construct prompts is to distinguish between what you want it to do and the conditions it must follow.
The task might be:
> “Turn this text into an explanation for people with no technical knowledge.”
The conditions might be:
Mentally separating these two elements helps you avoid forgetting an important part of the brief. It also makes it easier to review the prompt: first check whether you have requested a specific action, then check whether you have explained how it should be carried out.
An AI can generate a confident-sounding response even when it does not have enough information. If you need exact data, it is not enough to ask for “reliability.” You must establish what it should do when evidence or context is missing.
You can include instructions such as:
These instructions do not automatically make the response true. The AI can still make mistakes even when you ask it to be cautious. Their purpose is to reduce fabrications and make uncertainties visible so that you can check them.
DO NOT CONFUSE CONFIDENCE WITH ACCURACY
A well-written response may contain errors. A firm tone, an orderly structure, and specific details do not prove that the information is correct. Check figures, dates, legal references, medical information, proper names, and any claim that could have consequences separately.
Before you write, answer these six questions:
1. What do I want to achieve?
Define the main action: summarise, compare, explain, classify, write, review, or generate ideas.
2. Who is the result for?
Indicate the reader’s level of knowledge, role, approximate age, or needs when relevant.
3. What context should the AI know?
Add the information that changes the response.
4. What should it include and avoid?
Specify the required content and the limits.
5. How do I want to receive it?
Define the structure, length, and format.
6. What should it do if it does not have enough information?
Ask it to formulate questions, flag uncertainties, or avoid filling in the missing data on its own.
You can turn those answers into a template:
> Task: [specific action].
> Context: [necessary information].
> Audience: [person who will use the result].
> Requirements: [mandatory elements and limits].
> Format: [structure, length, and presentation].
> Uncertainties: [what it should do if information is missing].
You do not always have to copy these labels. They are a thinking aid. With practice, you will be able to write more natural prompts while keeping the same logic.
FROM A VAGUE REQUEST TO A PRECISE BRIEF
Initial request:
> “Explain how to organise my week better.”
Improved version:
> “Help me organise a Monday-to-Friday working week for a self-employed person who works from home. I have fixed meetings on Tuesday and Thursday mornings, need to set aside at least six hours for focused work, and want to leave some time each day to respond to emails. Propose a rough schedule in one-hour blocks. Do not include activities before 8:00 a.m. or after 6:00 p.m. Add three simple rules for adjusting the schedule when an urgent task comes up.”
The second request does not describe the person’s entire life. It provides only the information that affects the planning.
A long prompt can be confusing if it mixes instructions, examples, and data without any order. Precision does not depend on the number of words, but on whether the important information appears clearly.
A brief prompt can work very well:
> “Summarise this report for the management team in 120 words. Include the main problem, the two most important figures, and the recommended decision. Do not add information that does not appear in the report.”
By contrast, an extensive prompt can still be imprecise if it contains phrases such as “make it perfect,” “be very thorough,” or “surprise me” without explaining what they mean.
Start with the essential information. Then, if the response does not meet your needs, add a specific condition. Instead of writing a huge brief from the outset, you can work in two steps:
For example:
> “The structure is suitable, but I need the recommendations to be feasible with a budget of 200 euros per month. Rewrite only that section and explain the estimated cost of each action. If you cannot calculate it using the available data, state this.”
This turns the conversation into a review process. The first response does not have to solve everything if you know what you will check afterwards.
Before sending a prompt, look for the decisions you have left to the AI. Ask yourself:
If the answer to any of these questions is “no,” you probably still have an improvised question rather than a precise brief.
The goal is not to control every word the AI produces. It is to give it enough direction for the response to move towards your purpose from the first attempt. A good prompt reduces assumptions, makes your expectations visible, and allows you to correct the result with specific instructions. That is where the difference between asking for something and commissioning it begins.
An AI assistant can draft an email, summarize a document, suggest ideas, explain a concept, or transform a text. It does so based on the instructions and information available in the conversation. It does not “read your mind” or automatically know the outcome you expect.
This difference explains many failures. You might write:
> Give me ten ideas for improving my business.
To you, this might mean: “suggest specific, inexpensive actions that can be applied to a language academy with few adult students and barely any social media presence.” To the AI, it simply means: “generate ten suggestions related to improving a business.” The result will probably be generally correct, but of little use in your situation.
AI responds to the request it receives, not to all the details you had in mind.
An AI assistant works especially well when it has to produce or transform information under clear conditions. For example:
These capabilities do not work independently. Quality depends on the starting material and the conditions you specify. To summarize a text well, AI needs to receive the complete text or have reliable access to it. To compare two options, it needs to know the criteria that should be used. To draft an appropriate response, it needs to know who will receive it and what outcome you are seeking.
That is why saying that AI “knows how to write” is insufficient. It can produce a text that appears correct, but it will need instructions to decide what information to include, how much to develop each point, what tone to use, and what it must not invent.
EXAMPLE
A vague request would be:
> Write a post about productivity.
A more useful request would be:
> Write a 120- to 150-word LinkedIn post for self-employed people who work from home. Explain one simple way to plan the three most important tasks of the day. Use a direct tone, avoid motivational phrases, and end with a question.
The second request resolves several uncertainties: who will read the text, what topic it should cover, how long it should be, what tone it should have, and how it should end.
A response can sound confident and contain false information. This happens because the assistant generates a likely continuation based on linguistic patterns; it does not work like a database that automatically checks every sentence before displaying it.
It may invent:
These inventions are often called hallucinations, although the term can obscure the main problem: the response must be verified when accuracy matters.
It is not enough to ask it to “not invent anything.” That instruction may help guide the response, but it does not turn the assistant into a reliable source. You should ask it to distinguish between provided information, inferences, and aspects it cannot confirm. Then check important facts yourself using appropriate sources.
A practical instruction would be:
> Use only the information included in the text. If a fact does not appear, write “not stated” instead of filling it in with an assumption. Separate facts from interpretations.
Even so, that instruction does not replace review. It is a way to reduce risk, not a guarantee.
AI may interpret a brief request in several ways. If you write “make it more professional,” it does not know precisely whether you want:
The word professional describes a general intention, but it does not define a specific outcome. AI will choose an interpretation based on the available context and common patterns. It may get it right, but it may also change the text in a way you did not want.
When a word allows for several meanings, it is better to replace it with observable instructions. Instead of asking “make it more professional,” you can write:
> Keep the ideas and facts from the original. Use a formal but approachable tone, remove colloquial expressions, divide the text into short paragraphs, and avoid promotional adjectives.
These instructions make it possible to review the result against clear criteria. You no longer have to decide only whether it “sounds professional”; you can check whether it preserves the content, removes colloquialisms, and follows the requested structure.
Long requests often contain several tasks. For example:
> Summarize this report, extract its conclusions, identify the risks, prepare a table with the figures, and draft an email for the team.
The response may summarize the report well and forget the table. It may also include the risks but fail to draft the email. The more tasks you combine without ordering them, the harder it will be to check whether all of them have been completed.
A simple solution is to separate the assignment:
You can also request a specific output structure:
> Return five numbered sections with these exact headings: Summary, Conclusions, Risks, Figures, and Email.
The structure does not guarantee that the content is correct, but it makes omissions visible. If a section is missing, you will notice immediately.
An assistant may not know about recent events, legal changes, current prices, new versions of a program, or data published after its knowledge period. Even when it has access to the Internet or search tools, it may consult an incomplete source, misinterpret a page, or mix information from different dates.
For this reason, questions about current information require extra care:
There is a difference between asking:
> What is the legal limit?
and asking:
> According to the regulations in force in Spain in November 2026, what limit applies to this situation? State the regulation and the date of the source. If you cannot check whether it is still in force, say so explicitly.
The second request defines the scope and requires a possible lack of verification to be made visible. Even so, the final check remains the user’s responsibility.
IMPORTANT NOTICE
A clear, detailed, and confident response is not proof that it is true. Before using information that could cause harm, check its source, date, and context. This precaution is especially important for medical, legal, economic, employment-related, or security-related information.
If you ask the same question several times, you may receive different answers. The response may also change if you slightly modify the context, the model, the settings, or the order of the instructions.
This does not necessarily mean that one of the answers is correct. Variation is part of how these systems work. Therefore, if you need a repeatable result, you must define more conditions:
For example:
> Analyze each request according to these criteria, in this order: urgency, impact, and effort. Assign a score from 1 to 5 to each criterion. If information is missing, write “data pending” and do not calculate a final priority. Return one row per request with the following columns: request, urgency, impact, effort, and priority.
This assignment does not eliminate all variation, but it reduces implicit decisions. AI has less room to interpret for itself what it should do.
A good instruction does not try to turn AI into an absolute authority. It defines what work it must perform and sets limits so that you can review the result.
You can require it to:
You should pay particular attention to:
AI can save you time when preparing, organizing, and drafting. It cannot promise truth, perfect understanding, up-to-date information, professional judgment, or responsibility for the consequences. The aim of a good prompt is not to eliminate these limitations, but to make them visible and work with them in a controlled way.
Before giving an instruction, ask yourself two questions: What exact result do I need? and How will I check that the response is useful? The first defines the assignment. The second turns a vague impression into a review criterion.
The Minimal Skeleton of a Prompt That Works
A prompt does not have to be long to be useful. It has to contain the information the AI needs to understand the task and decide how to respond.
When you write:
> Give me ideas for improving my business.
the assistant has to guess too many things:
The AI may produce a grammatically correct response, but that does not mean it understood your intention. It will probably offer general recommendations that could apply to almost any business. It is not that it ignored your request. It responded to the limited information you gave it.
A basic prompt can be built from four parts:
You will not always need all four parts to the same extent. For a simple task, one sentence may be enough. For important work, it is worth developing each section.
The task indicates the main action. It is the part that answers the question: what do you want the AI to do?
It is best to use precise verbs:
There is an important difference between asking for “help with a text” and asking to “summarize this text in five points for someone who is unfamiliar with the subject.” In the second case, the action is defined and the AI has fewer decisions to make on its own.
A specific task can also include the object of the work:
> Summarize the following report.
> Rewrite this email.
> Suggest ten titles for this article.
> Extract the dates and names of organizations mentioned in the document.
Avoid piling up several different tasks without any order. If you write, “analyze this report, summarize it, correct the style, look for errors, and prepare a presentation,” the AI may try to do everything superficially. When the work is extensive, break it into steps or indicate which task has priority.
Context is everything the AI needs to know to avoid responding in a generic way. It does not mean telling it your entire story. It means including the details that affect the result.
Notice the difference:
> Write a social media post about my course.
compared with:
> Write a LinkedIn post promoting an Excel course for adults who need to organize their administrative tasks more effectively. The course is online, lasts four weeks, and starts on October 15. The goal is to attract registrations, using a clear and professional tone.
The second request provides information about:
You do not need to include all these details in every prompt. The practical rule is simple: add any information whose absence could significantly change the response.
To check whether context is missing, ask yourself:
Context can also include an example. If you want the AI to imitate a structure, length, or level of formality, show it one. Explaining “write in a friendly way” may be less useful than including a paragraph that represents the tone you are looking for.
Saying “do it well” is not a useful instruction. The AI does not know what you understand by a good response. You must describe the result in a way that can be checked.
You can specify aspects such as:
For example:
> Write an email of between 120 and 150 words, in Spanish, with a professional and cordial tone. Include a brief explanation of the delay, a new delivery date, and a request for confirmation.
This instruction is more useful than “write a professional email about the delay.” The first reduces the number of decisions the AI has to make.
You can also request a specific format:
> Organize the response in a table with these columns: problem, probable cause, and recommended action.
> Present a five-line summary first and then expand on each point with an example.
> Return a numbered list of eight options. For each option, include the title, the main advantage, and a possible difficulty.
Format is not a decoration. It helps make the response easier to review, copy, or reuse.
Conditions indicate what the AI must follow while carrying out the task. They are especially useful when you want to avoid common errors.
Some examples are:
> Do not invent figures or sources. If a piece of information is missing, say so.
> Use only the information included in the text.
> Avoid technical terms and explain each concept as you would to a beginner.
> Do not recommend paid tools.
> Keep proper names and dates exactly as they appear.
> If there are several possible interpretations, separate them and explain what information is missing to choose between them.
These instructions do not automatically make the response accurate. An AI may state something incorrect even if you ask it not to invent information. However, conditions make the criterion it must follow more explicit and make subsequent review easier.
IMPORTANT NOTICE
Asking the AI “not to invent information” is not a substitute for checking the information. When the response includes data, laws, prices, dates, references, or important claims, verify those elements with appropriate sources.
These four parts can be combined into a simple structure:
> Task: [what the AI must do].
> Context: [necessary information about the situation, the material, and the audience].
> Result: [what it must provide, with what length and in what format].
> Conditions: [limits, criteria, and information it must handle carefully].
You do not have to write the labels “Task,” “Context,” “Result,” and “Conditions” literally. You can do so if it helps you organize the request, especially when the task is long.
For example:
> Summarize the report below. It is intended for small-business managers without a financial background who need to decide whether they can take on a new monthly expense. Explain the main revenues, costs, and risks in simple language. Present a five-point summary first and then a conclusion of no more than 100 words. Use only the data in the report; if information is missing to draw a conclusion, point it out.
This prompt contains:
Vague request:
> Give me ideas for organizing my week better.
Improved request:
> Suggest ten ways to organize my workweek. I work as a self-employed professional, see clients in the mornings, and set aside the afternoons for administrative tasks and my own projects. I often have unexpected meetings and find it difficult to estimate how much time I need for each assignment. Present the ideas in a table with these columns: method, how to apply it, and possible drawback. Prioritize free solutions that can be managed with a digital calendar. Do not include specific apps.
The second version does not guarantee that all the ideas will be good, but it establishes a clear framework. The AI knows what problem it must address, in what situation, and within what limits.
You can also ask it to help you identify the missing information:
> I want to ask an AI to help me organize my workweek. Before suggesting solutions, ask me up to five questions about my schedule, tasks, and priorities. Do not give recommendations yet.
This strategy is useful when you are still not sure what the task should be. A good prompt does not always ask directly for the final answer. Sometimes it starts by gathering the necessary context.
Many prompts use formulas such as:
> Act as a marketing expert.
or:
> You are a patient math teacher.
Assigning a role can guide the tone and type of response, but it does not replace specific instructions. “Act as an expert” does not indicate what task it must perform, for whom, or according to what criteria.
It is more useful to combine the role with the task:
> Act as a secondary-school math teacher. Explain how to calculate a percentage to an adult who has forgotten the basic operations. Use an example involving a discount and end with three practice exercises, without including the solutions.
In this case, the role helps establish the level and style, but the task, context, and result remain the main parts.
You can copy this template and complete it whenever you prepare a request:
> I need you to [specific verb and task].
>
> Context: [explain the situation, objective, audience, and relevant information].
>
> Deliverable: [indicate the format, length, structure, and level of detail].
>
> Conditions: [state the limits, tone, what it should avoid, and how it should act if information is missing].
A completed example:
> I need you to revise this email to make it clearer and more cordial.
>
> Context: I will send it to a client who is waiting for a business proposal. The proposal will be delayed by three days because a supplier’s price has not yet been confirmed.
>
> Deliverable: Return a version of between 100 and 130 words. Keep the essential information and end with a request for confirmation.
>
> Conditions: Use a professional tone, do not overstate the apology, and do not invent a date other than the one provided.
Before pressing “send,” review these questions:
You do not need to turn every request into a lengthy document. For a simple task, one or two well-constructed sentences may be enough. The structure helps you think through the task; it is not there to add words for their own sake.
BRIEF EXERCISE
Turn this vague request into a complete prompt:
> Explain how to save more money.
Include a specific task, the person’s context, the response format, and at least one condition. For example, you can indicate their approximate income, usual expenses, goal, and available timeframe. Then check whether the AI would have to invent any information to respond.
The minimum skeleton does not guarantee a perfect response. It does achieve something more practical: it makes your expectations visible and reduces the decisions the AI has to make on its own. From there, you can add examples, evaluation criteria, or review steps when the task requires them.
Want to know what happens next? Get the full book:
Prices checked in the stores on 15 Nov 2026; they may change.
Back to the book page