Welcome! This material was born digital on purpose: like the news and discoveries of the field, it can be reviewed and updated at almost the same speed — something impossible in a printed book.

The goal is not to exhaust the subject, but to serve as apractical reference guide: a place from which you extract a specific piece of information, copy a prompt and adapt it to your needs. No need to read in order — go straight to what you need.

Keeping theoretical definitions to the minimum needed: understanding how AI tools work is fundamental to extracting their best results. That is the only reason it makes sense to grasp the basics of how these tools work — and it will not be too complicated. As you have probably heard, ChatGPT (and related systems) works like a huge text predictor. We are more or less used to these features in some everyday apps. For example, on a phone, when we reply to our e-mails, the message barely starts and a suggestion for the following text already appears (Figure 01).

That autocomplete happens because of “context” and because of the probability of words appearing in sequence. If I ask you to complete the sentence “the sky is [___]”, which word came to your mind? The probability that you thought of “blue” or “big” is much higher than that of having thought of “crimson”. The sentences “The sky is blue”, “The sky is big” and “The sky is crimson” are all technically correct, but they are not used with the same frequency in our daily lives.

How does that apply to AI tools? They were presented ("trained") by analyzing billions — literally — of texts and materials, so they hold data on the probabilities of words occurring in sequence. Obviously it is far more complex than that, but this analogy will help us use the tools more effectively.

In the context of this e-book: treinamento, quite simply, is the name given to the process of presenting sources of material (for example, texts in different languages) to the AIs, so that patterns can be recognized and, from there, the decision can be made about which text or word should come next.

When you open AI tools (such as Gemini and ChatGPT), you will end up facing the possibility of choosing "different" AIs — what we call models. Think of them as "versions" of the AIs, in the same way that we have bicycles (no gears, 18 gears, suspension, carbon fiber...) or cars (Onix LT, LTZ, Premier...): it is more or less the same logic. There are no "right or wrong" models: there are models better suited to each task, with different capabilities and prices (or usage limits).

Two important categories come in here: models called reasoning (thinking models) and "common" models. In theory, reasoning models are better suited to complex tasks, which involve considering multiple factors. They are more capable in terms of processing, but also slower to generate the answers.

In practice, do not worry too much about this: currently (2026) practically all relevant models are reasoning models by default, and what changes is how much they "think" before answering — many tools already offer a reasoning-intensity selector (something like "fast" versus "think harder"), instead of two entirely separate products. For simple tasks (summarizing, translating, formatting), prefer the fastest option; for analyses that require cross-referencing multiple pieces of information, it is worth turning on extended reasoning and waiting a little longer. The idea remains the same: test, without fear, the available products and models and check which ones give the best answers for each task.

A frequent question I get is: "which is the best tool?" The best answer will always be "it depends", but let me elaborate. Which is better: a hammer, a screwdriver or a pair of pliers? It depends on what you need to do.

With AI assistants, the logic is similar. GPT, Gemini and Claude are the best known. GPT has the best Deep Research available (advanced search/research mode — to be explained later). Gemini integrates with the Google ecosystem and has improved considerably since its creation. Claude, in turn, has excellent results for text production and for tasks involving programming.

If you are considering paying for an AI assistant, follow these steps:

1. Test the free versions of the services you are considering for at least one to two months.
2. Do not subscribe to the so-called "AI aggregators" (sites that offer access to several models at low/promotional prices): the most likely outcome is that they will not meet your needs properly. With the more affordable Gemini and ChatGPT plans available, it is not worth using intermediaries.
3. You can use Google's models, free of charge, at https://ai.dev.

Below are listed the main AI tools available, free or paid.

Let us start with the prompt. This is a fundamental concept: prompt is the name given to the instructions we pass to the artificial intelligence tools.

"The text below will be submitted as an abstract for a conference paper in [psychology/engineering/law/medicine]. Review the grammar, coherence and cohesion of the text, suggesting changes whenever needed."

"An electrician came to my home and said I need to install a residual-current device (RCD). Explain what it is and what its function is."

"You will act as an expert radiology instructor. I have 20 days to learn the fundamentals of ultrasonography and you will be my teacher: build a detailed learning schedule for those 20 days, covering the key ultrasound concepts that an excellent introductory module should include."

"You will act as an English teacher specialized in teaching English as a foreign language. I have 20 minutes per day and you will be my instructor: build a detailed learning plan with basic, intermediate and advanced English content so that a foreign learner can communicate adequately.
                
We can begin immediately, and your first task is to assess my current English level. Ask the necessary questions to evaluate my skills."

You are a senior career coach with 20 years of experience in professional transitions, specialized in helping IT, engineering and creative professionals change careers with clarity and confidence. Your style is empathetic, direct and evidence-based, inspired by methods such as Ikigai and GROW.

**Context**: The user is considering a career transition — changing jobs, industries or even starting a business — but needs to reflect deeply to avoid regrets. We are in an iterative conversation to map the ideal path.

**Objective**: Guide the user to reflect on their current moment, identify real motivations, strengths and obstacles, creating a personalized transition plan in clear, actionable steps.

**Tasks (reverse prompting)**: Do NOT give advice or plans yet. Instead, ask 5-7 essential, open questions to gather key information. Group them into categories (e.g., "About your current situation", "Motivations and values", "Skills and network"). After the user's answers, use them to generate a reflection report and an initial plan.

Initial questions to ask NOW:
1. What is your current profession, how long have you been in it, and what do you like/dislike most about it?
2. Why do you want to make a transition now? Describe the emotional or practical triggers.
3. What are your main skills (technical and soft skills) and the achievements you are proudest of?
4. What would "success" in the new career mean for you (e.g., salary, flexibility, impact)?
5. What real obstacles do you foresee (finances, family, market)?
6. Describe your professional network and who could help in this transition.
7. On a scale of 1-10, how ready do you feel to make the change in the next 6-12 months?

Analyze the answers and confirm understanding before proceeding.

[CONTEXT] We are in a long interaction about [INSERT TOPIC, e.g., the development of a protocol]. To make sure we do not lose important details and to clear the context window in a new chat, I need you to condense everything we have done.

[TASK] Generate a detailed "Current State Summary" containing:
1. The main goal we are pursuing.
2. All decisions already made and validated (what is already done).
3. The exact current status of where we stopped.
4. The immediate next steps that were pending.
5. The style constraints or preferences I have already taught you in this conversation.

[FORMAT] The output must be structured text, ready to be copied and pasted as the FIRST prompt of a new chat, so the new AI instance can continue the work exactly where we left off, without losing context.
                    

From the discussion of the previous section, we understand that some prompts are more or less suited to a specific task. From there, a frequent question arises: is there a "perfect" prompt? The answer is: there is no single perfect one, but there are components that, when well organized, produce excellent results for most of the tasks we need to solve.

Elements of a well-built prompt

1. Persona — who the AI should "embody" when answering.
2. Objective — what you want to achieve with that output.
3. Contexto — the background information that makes the task specific.
4. Task — the concrete command, what the AI should do.
5. Constraints — what should (or should not) appear in the response.
6. Output format — how the result should be structured (table, list, running text...).

Note that not all components are mandatory. For example, for a simple task, such as generating a text, you may not need persona, context or constraints.

Example 1

Persona:
You will act as a human-resources manager.

Objective:
Create objective criteria for evaluating administrative performance.

Context:
An administrative team composed of assistants and analysts.

Task:
1. Define measurable criteria.
2. Propose an evaluation scale.
3. Suggest an assessment frequency.

Response format:
A table with criteria, description and scoring scale.

Example 2

Persona:
You will act as a corporate-communication consultant.

Objective:
Draft an internal communication about the implementation of a new administrative system.

Context:
Staff with different levels of technology familiarity.

Task:
1. Explain the reason for the change.
2. Highlight operational benefits.
3. Inform about next steps.

Constraints:
Use simple, accessible language.
Avoid jargon and technical terms.

Response format:
An institutional announcement in continuous text.

Example 3

Persona:
You will act as the internal compliance officer.

Objective:
Assess administrative risks in a vendor contracting process.

Context:
Recurring hiring of outsourced services.

Task:
1. Identify operational risks.
2. Identify basic legal risks.
3. Propose simple administrative controls.

Response format:
A table followed by a brief conclusion.

From simple to complex: three levels of the same prompt

In general terms, prompts can be more or less extensive and can contain more or fewer elements. There are contexts in which simple, direct prompts will work very well, and in others you will need prompts with far more detail and elements. See below a few more examples:

Basic

You are a text editor. Review the abstract below for grammar, cohesion and coherence. Structure it into paragraphs.

Intermediate

You are a scientific editor specialized in medicine with 15 years of experience in international journals.

This abstract will be submitted to JAMA as a structured abstract.

Review the abstract below for:
- Grammar and clarity in scientific English
- IMRAD structure (Introduction, Methods, Results, Discussion)
- A maximum length of 300 words

Expected format: paragraphs with bold subsection headings.

Advanced

[SYSTEM]
You are a senior scientific editor specialized in medicine with 15 years of experience in international journals (JAMA, Lancet, NEJM). Your task is to perform a critical editorial review.

[CONTEXT]
This abstract will be submitted to JAMA. Audience: editors, reviewers and international clinical readers. Expected standard: editorial excellence.

[ACTION]
Review and rewrite the abstract below while preserving scientific integrity and maximizing clarity.

[POSITIVE CONSTRAINTS]
- Structure: IMRAD with bold subsection headings
- Length: maximum 300 words
- Style: formal scientific English, active voice whenever possible
- Citations: use bracketed numbers [1], [2], etc.
- Data: preserve all original numbers and values exactly

[NEGATIVE CONSTRAINTS]
- Do not speculate beyond the presented data
- Do not use nonstandard jargon or undefined abbreviations
- Do not alter the scientific conclusions; only clarify expression

[EXPECTED FORMAT EXAMPLE]
**Introduction.** [2-3 sentences on the problem and the knowledge gap]
**Methods.** [Design, population, intervention]
**Results.** [Main findings with values]
**Discussion.** [Clinical meaning, limitations, next steps]

[VALIDATION]
Before answering, confirm:
✓ Does every claim have a supporting source or citation?
✓ Is the length within 300 words?
✓ Is the IMRAD structure preserved?

If any criterion is not met, state which one before delivering the revision.

Advanced level: quality checklist

To evaluate whether your prompt is "good enough", use this checklist:

"Answer about the effectiveness of vitamin D in COVID-19. You MUST provide answers based only on scientific evidence. Cite your sources explicitly."

Although they are semantically identical, Version A tends to produce better compliance because the model reads the restriction first and treats it as a high-priority policy for the task.

Recency and primacy

Language models tend to give more weight to information at the beginning and at the end of the prompt:

- Instructions at the beginning influence the overall response policy. If you define tone, persona, or foundational constraints upfront, the model tends to treat them as guiding principles (primacy effect).

- Instructions at the end influence the immediate action. If the final instruction is "answer in 100 words," the model will usually prioritize that rule in the current output, even when earlier parts pull in another direction (recency effect).

This leads to a practical problem: instructions buried in the middle of a long prompt are often the easiest to lose. They usually have the lowest compliance rate.

Example:
Constraints on sources (PubMed only, last 5 years) show lower compliance when placed in the middle than when placed right after the persona — this small adjustment can mean significantly better performance.

Example:

Start: "You MUST cite only primary sources."
End: "Before finishing, confirm that ALL claims include a citation from a primary source."

Prompt chunking

In certain situations, you may have to write a necessarily long prompt. One of the best techniques for running a long prompt with good results is to split it into blocks/sections and use specific formats (markup, JSON). Instead of a 2,000-word prompt with several instructions scattered around, create:

BLOCK 1: [Persona + core constraints]
BLOCK 2: [Context + problem]
BLOCK 3: [Action + expected format + validation]

The finishing touch
There are more advanced techniques for output checking (covered later in this e-book), but one simple command can already help in factual tasks: "Review the output before reproducing it, and if it does not match the requested input, restart the process. Continue until there are no inconsistencies or incoherences."

Example: "[Insert your system prompt here] To confirm that you read these instructions, begin your answer with: 'I understand that I must [brief summary of the critical instruction].' Then continue with the response."

📐 Recommended order for professional prompts

It is possible to customize the GPT according to everyday use and make it even better suited to our needs. Go to the ChatGPT settings and go to "personalization":

General "master" prompt

By default, I expect responses to be formal, concise and well structured. Most of my requests relate to producing scientific documents, directly or indirectly.

Do not automatically agree or disagree with me. Analyze my prompts and respond as objectively as possible.

To ensure these guidelines are actually followed, I will always start my interactions with "let's work" or "vamos trabalhar" and expect to be addressed as ["[YOUR NAME]"]. In those situations, responses must be absolutely neutral: do not seek agreement or disagreement with my questions, thoughts or opinions. Critically analyze the text provided and present the most appropriate answer, strictly avoiding errors, inaccuracies and hallucinations.

Projects: custom workspaces

OpenAI itself defines Projects as specialized "workspaces". Imagine that you have multiple roles in the same company or across different companies: you can set up one Project for each, with instructions and files customized for each context.

Each project allows a master prompt associated with that activity, plus reference files. The combination of a precise prompt with relevant documents transforms projects into true personalized assistants — and very useful ones. By the way: Gemini calls them "Gems"; Claude calls them "Projects". Most AIs have some similar kind of customization.

Prompt example for a language-learning Project

You will act as a teacher and consultant specialized in language teaching for adults, grounded in contemporary methodologies such as Task-Based Learning, Communicative Language Teaching and related approaches.

You will use andragogy and neuroeducation to design instruction programs in any language, prioritizing communicative competence and functional fluency before grammatical precision.

Tasks include:
- Designing structured programs in a modular architecture, adapted to the learner's CEFR level (A1-C2) and specific goals
- Conducting practical conversation sessions with progressive immersion
- Producing exercises anchored in spaced repetition, interleaving and retrieval practice
- Immediate, specific, action-oriented feedback
- Continuous formative assessment with dynamic difficulty adjustment

Difference between Projects and GPTs

What is today called "GPT" inside ChatGPT used to be called "plugin" and was considerably different from Projects. As both tools advanced, they became increasingly similar. In practical terms: understand GPTs as personalized "work environments" made to be shared with other people, while Projects are made for individual use, even though they can also be shared.

The standard search (Google or a common AI) is like asking a librarian about a subject and having them hand you the most recent book they read. You get fast, but limited, information. The Deep Researchis like hiring a professional researcher: he (1) creates a research plan, (2) reads dozens of articles/sources, (3) cross-checks contradictory information, (4) looks for primary sources to break the tie and (5) writes a consolidated report. It takes longer — from a few minutes to half an hour —, with an incomparably deeper result.

Deep Research is useful at both extremes of knowledge: when you "know nothing" (to build a minimum repertoire before a meeting with a specialist) and when you "know a lot" (to update yourself quickly on the latest in the literature). Specify the sources Deep Research should prioritize: full-text articles on PubMed, case law from a specific court, clinical practice guidelines from the Ministry of Health, etc.

You will conduct a deep research review on mental health and the use of LLMs (large language models).

Provide examples or cases, risks, global and Brazilian prevalence of the most frequent mental disorders, and discuss whether the use of LLMs may or may not represent a threat to the general population.

Also provide journalistic coverage of the topic and explain how major technology companies are responding to it.

In addition, search for and analyze scientific papers related to the topic (EBM criteria with A/B level evidence, only papers with full text available on PubMed/Medline) and summarize their main findings.

Comparison: Deep Research vs. Deep Search

Summary of the choice: Need a technical dossier? ChatGPT ou Gemini. Need to check facts and news? Perplexity ou Grok. Need to analyze 50 PDFs you already have? Claude ou Gemini Notebook (which we will see next).

Note: ZCode, from Z.ai (listed in the agentic tools grid), is not a research tool — it is an agentic development environment for programming with the GLM models, equivalent to Claude Code or Codex. That is why it does not appear in this Deep Research table.

The Canvas mode allows you to edit and review the content being built. Let us work with a practical example: you need to review a document you created (minutes, an article, a book chapter...) and, instead of simply asking the AI to edit it for you, you can open the document in Canvas and edit it together with the AI, directly editing what you understand should be edited and asking the AI to make other edits (summarize a paragraph, expand a section...).

Pay attention to a recent change: ChatGPT removeu the Canvas mode from its main models in 2026 — text and code editing now happens directly inside the conversation itself, in editable "blocks", without a separate panel. Today, Gemini is the reference tool for this type of dedicated collaborative editing space, and it is where the example below works best.

Canvas can be an extremely powerful tool for editing code, pages (HTML) and even documents. I prefer to edit my own texts and I will share, in the field below, the prompt I use so that you can copy it and adapt it to your reality:

You will review the attached text for possible grammar and typing mistakes. Evaluate its coherence and cohesion, but do not edit it directly: use strikethrough formatting on the portions you believe should be removed and, beside them, add your suggestions in bold, so I can identify the parts that should be revised and judge them myself.
            

Worth remembering that this kind of feature changes fast — GPT's own Canvas existed and was discontinued within the lifetime of this e-book. The best way to understand and learn the tool's uses is by practicing, so let us get to work!

In July 2026, Google renamed NotebookLM to Gemini Notebook, as part of the closer integration with the rest of the Gemini ecosystem. It is the same tool, with the same purpose — if you used NotebookLM, your notebooks remain exactly where they were, and old links keep working. In this e-book, I use "Gemini Notebook" from now on, but do not be surprised if you still find the old name around.

It is my favorite AI tool, and I will explain the reasons. Some time ago, it became famous for the possibility of generating "audio summaries" — which, despite being an excellent feature, is not the main reason I use it so much.

Let us revisit the conversation about hallucinations: through adjustments to the prompt and with the anchoring (grounding) of the sources, we can reduce the hallucination rate to a minimum. That is exactly what Gemini Notebook does: it only consults — and answers — from the sources you choose to use.

When you create a notebook, the first screen is for choosing sources. You can add documents, Drive files, Google Docs (including spreadsheets), YouTube videos, website URLs and audio files. All this material becomes your source, and todas the answers extracted from your prompts will come exclusively from the sources you added. If you have already heard of RAG (retrieval augmented generation), that is more or less what Gemini Notebook does — reducing errors and hallucinations to the minimum possible.

How to add sources in Gemini Notebook

Once the sources are added, the interface is organized as follows: on the left, the added sources (you can remove them or add others); in the center, the prompt interface; on the right, the interaction resources that Gemini Notebook provides.

Resources available in Gemini Notebook

The resources panel includes:

• Audio summary (a podcast with two hosts discussing your sources)
• Study guide (with multiple-choice, short-answer, and essay questions)
• FAQ (frequently asked questions about the content)
• Summary and timeline
• Briefing document (executive summary)
• Mind map
• Infographics
• Flashcards and quizzes
• Data analysis with code (most recent feature: a secure cloud environment that writes and runs code to cross-reference your sources, generating graphs, spreadsheets and reports from them)

Almost all the resources show a "pencil" icon — which means you can edit the default instructions of that feature by inserting a custom prompt.

Editing resources — audio example

You will produce an audio file that is as detailed and extensive as possible, addressing methodological details with rigor and precision: describe in depth how the structure of the developed product was set up, the details of the debate prompt (whether it involved an agentic strategy with multiple personas, or whether different agents were in fact used). Finally, dwell on the results, describing them carefully and explaining their implications for health care.

Large context window

Gemini Notebook offers a 1-million-token context window, allowing interaction with multiple extensive documents simultaneously — entire books, scientific articles, meeting transcripts. Unlike conventional AIs, it was designed specifically to "consult professionally" multiple files: its sources stay separate from the conversation window but fully accessible to the model. This allows you to converse at length with the documents without their size disturbing the conversation's memory.

Each answer comes with the precise citation of the original excerpt of your document, making verification easier and dramatically increasing reliability for professional use.

Google ecosystem: integration and dynamic updating

Gemini Notebook integrates with the Google ecosystem, allowing you to use Google Drive folders and documents as sources. The most useful point: when you connect a Google Docs as a source, Gemini Notebook keeps the link with the original file — which you can update whenever you want. When it detects that the latest version has changed, Gemini Notebook provides a button to update the source.

Audio and video transcription

Gemini Notebook allows you to fully transcribe audios and videos: just add them as sources and request, via prompt, the transcription of the material. You can include highlighting instructions: imagine you recorded a meeting that will need to be turned into minutes, or a class where the teacher frequently highlights important information. Examples:

Transcribe the full source material, identifying, whenever possible, the different speakers.

Then draft formal meeting minutes including the following elements: date, identified participants, agenda discussed, decisions taken, and follow-up items with the person responsible for each.

Gemini Notebook free vs. paid

There are differences between the free and paid versions, changing the number of notebooks that can be created, the number of sources per notebook, the number of resources created/day and, subjectively, the speed at which some resources are generated (personal impression). My suggestion: start working with the free version and only after becoming familiar with the tool migrate to the paid version (if needed). The paid tier is not sold separately: it comes included in the Google AI Plus, Pro and Ultra subscriptions.

• More notebooks and more sources per notebook (higher limits than the free version)
• More audio-summary generations per day
• Longer, more detailed audio outputs
• Deeper integration with Google Workspace
• Notebook-sharing features for teams
• Priority support and early access to new features

Gemini Notebook is especially useful in academic contexts! I use it a lot to interact with academic texts (PDF articles, for example — and now EPUB files too)

• Literature review: add the selected articles and use the prompt to synthesize the main findings, compare methodologies or identify knowledge gaps.
• Guideline study: load the most recent guidelines of a specialty and ask clinical questions directly to them
• Presentation preparation: use the Study Guide and FAQ features to generate questions about the content of classes or lectures
• Learning diary: keep a Google Docs with daily notes and connect it to Gemini Notebook; at the end of the week, generate a summary or quiz about what you learned.
• "Pro" tip: use Gemini Notebook as a personal tutor! Study an article and explain it to the tool, asking it to correct and complement any critical points you got wrong or missed.

What is OpenEvidence?

OpenEvidence is an AI-based clinical decision support platform, developed specifically for physicians and health professionals. Unlike general-purpose models (such as ChatGPT and Gemini), OpenEvidence was trained and optimized for medical queries, with access to specific medical literature curated by its editorial team.

The platform has content agreements with several reference organizations such as NEJM (New England Journal of Medicine), JAMA, NCCN (National Comprehensive Cancer Network), Cochrane, Wiley, ACC (American College of Cardiology), ADA (American Diabetes Association) and other medical societies.

How does it work?

The operation is simple and very similar to the logic behind the other LLMs:

1. You ask a clinical question in English or Portuguese
2. The tool searches its internal medical-literature sources
3. It returns an answer with citations and links to the original articles
4. You can ask follow-up questions in the same session to go deeper

Access it at: www.openevidence.com

Best prompt practices for OpenEvidence

As with any AI tool, the quality of the answer depends on the quality of the question. For OpenEvidence, the practices below make a significant difference:

1. Include relevant clinical context about the patient

Instead of asking generically "what is the treatment for hypertension?", provide context:

68-year-old male with a 10-year history of hypertension, type 2 diabetes, and an eGFR of 45 mL/min/1.73m². He is currently taking losartan 50 mg/day and metformin 500 mg twice daily. Current blood pressure: 155/95 mmHg. According to major guidelines, what therapeutic options are recommended to optimize blood-pressure control in this scenario?

2. Specify the type of information you want

What are the current indications for anticoagulation in non-valvular atrial fibrillation? Include:
- CHA2DS2-VASc score and treatment thresholds
- HAS-BLED score and its clinical usefulness
- Comparison between direct oral anticoagulants (DOACs) and warfarin in the main populations
- Special scenarios: advanced CKD, obesity, and reversal of anticoagulation

3. Request evidence level and study quality

What is the current evidence on the use of ivermectin for treating COVID-19? Please:
- Cite only randomized clinical trials and meta-analyses
- State the quality of evidence level (GRADE when available)
- Distinguish primary clinical outcomes (mortality, hospitalization) from secondary outcomes

4. Compare therapeutic options directly

Compare the efficacy and safety of SGLT2 inhibitors versus GLP-1 receptor agonists in patients with type 2 diabetes and established cardiovascular disease, based on randomized clinical trials with hard cardiovascular outcomes (MACE, cardiovascular mortality, hospitalization for heart failure).

5. Use follow-up questions to deepen the answer

OpenEvidence allows follow-up questions in the same session. For example, after obtaining an answer about initial treatment, you can ask:

What if this patient also has heart failure with reduced ejection fraction (HFrEF)? How does that change the therapeutic choices, and which combinations have the strongest evidence for mortality benefit?

Core cautions when using OpenEvidence

OpenEvidence is a powerful tool, but its inadequate use can lead to clinical errors. Three cautions are especially important:

Although AIs are powerful tools, their use demands responsibility and knowledge of their limitations, especially in critical areas such as health. Below we highlight four fundamental points of attention.

1) Hallucination (Hallucination)

The term "hallucination" in AI refers to the generation of information that seems correct and is presented with high confidence, but is false or invented. This happens because the model predicts the most probable next word, not necessarily the truth. In health care, this can mean the invention of clinical conduct, drug doses or citations of articles that do not exist. Always verify the sources.

2) Risk of bias (Bias)

AI models are trained on vast datasets from the internet, which contain society's prejudices and biases. This can result in responses that perpetuate gender, race or socioeconomic stereotypes, or that show bias in medical diagnoses (for example, underdiagnosing certain conditions in specific demographic groups). Be alert to these distortions.

3) The "black box" paradigm (Black Box)

We do not know exactly como or por que an AI reached a given conclusion. We do not know which materials were used in the training nor which specific sources underpinned a particular answer. Moreover, the internal processes of neural networks are complex and not very transparent. This is critical in clinical practice, where explainability is fundamental to the safety of and confidence in decision-making.

4) What the Brazilian regulation says

Since August 26, 2026, the use of artificial intelligence in medical practice in Brazil has been regulated by the CFM Resolution No. 2,454/2026. You do not need to memorize the details, but it is worth knowing the central principles, because they confirm — now also as a rule — exactly what we had been reinforcing in this e-book:

As with any recent rule, there may be updates or complementary regulations after the publication of this material — it is worth consulting the current text directly on the CFM website (cfm.org.br) before basing any compliance decision on it.

This section works as a reference catalog: more sophisticated prompting techniques, each with a generic example (ready to edit) and, in most cases, a second example already applied to a context — clinical, legal or administrative — so you can see the pattern in action. There is no need to memorize the English names; come back here whenever you need to copy and adapt one of these templates to your current task.

Prompt for decomposing complex tasks (for productivity)

[TASK DECOMPOSITION - REAL EXAMPLE]

My main task is: Prepare a 2025 departmental spending-analysis report for presentation to the executive board.

Please break this down into sequential subtasks, numbered from 1 to N.

For EACH subtask, provide:
- Subtask number and title
- Description of what needs to be done
- Success criterion
- Any prior dependency
- Estimated time

Additional context:
- We have data from three sources: SAP, internal spreadsheets, and department receipts
- The presentation is in 2 weeks
- Audience: 8 non-technical executives
- Expected format: 15-20 slides with the main charts

Editable example

[TASK DECOMPOSITION]

My main task is: [DESCRIBE THE OVERALL TASK]

Please break it into sequential subtasks, numbered from 1 to N.

For EACH subtask, provide:
- Subtask number and title
- Description of what must be done
- Success criterion (how to know it was done well)
- Any prior dependency
- Estimated time

Then present a recommended execution timeline.

Chain of Thought

Editable example:

[Chain of Thought - EDITABLE]
You are a [SPECIALIST IN WHAT?].
Analyze: [DESCRIBE THE CASE / PROBLEM]

Think step by step:
1. [WHAT IS THE FIRST LOGICAL STEP?]
2. [WHAT IS THE SECOND STEP?]
3. [WHAT IS THE THIRD STEP?]

Respond by detailing each step of your reasoning.

Complete example:

[Chain of Thought - FULL]
You are a project manager. Analyze the following: an ERP implementation project was planned for 6 months with a budget of R$ 500,000. We are now in month 4, R$ 450,000 has already been spent, but only 60% of the functionality has been delivered. The client is dissatisfied.

Think step by step:
1. Diagnosis: what was the root cause of the delay (scope, resources, planning)?
2. Impact analysis: how much extra time and cost will be required?
3. Strategy: what options exist (extension, scope reduction, larger team)?

Respond by detailing each stage.

Tree of Thought

Editable example:

[Tree of Thought - EDITABLE]
Analyze: [DESCRIBE A COMPLEX SCENARIO WITH MULTIPLE OPTIONS]

Build a tree of thoughts simulating multiple scenarios:

Scenario 1: [FIRST OPTION]
- Sub-consequence 1a: [...]
- Sub-consequence 1b: [...]
- Likely outcome: [...]

Scenario 2: [SECOND OPTION]
- Sub-consequence 2a: [...]
- Sub-consequence 2b: [...]
- Likely outcome: [...]

Scenario 3: [THIRD OPTION]
- Sub-consequence 3a: [...]
- Sub-consequence 3b: [...]
- Likely outcome: [...]

Synthesis: Which scenario has the highest probability of success?

Complete example:

[Tree of Thought - COMPLETE]
A company needs to decide about international expansion. Analyze the three main scenarios:

Scenario 1: Asian market (high risk, high return)
- Initial investment: $10M
- Regulatory barriers: high
- Market potential: 500M people
- Likely outcome: Return in 4-5 years if entry is achieved

Scenario 2: Latin American market (medium risk, medium return)
- Initial investment: $5M
- Regulatory barriers: medium
- Market potential: 150M people
- Likely outcome: Return in 2-3 years

Scenario 3: European market (low risk, low return)
- Initial investment: $3M
- Regulatory barriers: high
- Market potential: 100M people
- Likely outcome: Stable return in 1-2 years

Synthesis: Latin America offers the best risk-return balance for short-term expansion.

Chain of Learning

Editable example:

[Chain of Learning - Editable example]
You will act as an interactive simulator for my training in [WHICH DOMAIN?].

[Your role]
1. Create an initial scenario or problem about [SPECIFIC TOPIC]
2. Do not reveal the solution immediately
3. Provide neutral information as I progress
4. Adjust difficulty according to my performance
5. Review my reasoning and point out gaps

[Interaction format]
- I present my analysis or solution
- You provide constructive feedback
- We iterate until I reach mastery

I am ready to begin. What is the domain and topic?

Complete example:

[Chain of Learning - Complete example]
You will act as an interactive simulator for my training in Clinical Epidemiology.

[Your role as teacher-simulator]
1. Create a realistic outbreak scenario (for example: "We detected 15 cases of gastroenteritis in a day care center over 3 days")
2. Do not reveal the most likely etiology immediately
3. Provide data as I request it (symptoms, incubation, exposures, laboratory findings)
4. Increase complexity as I demonstrate understanding (for example: confounding factors)
5. Review my epidemiologic reasoning (was my initial hypothesis appropriate?)

[My approach]
- I will formulate hypotheses about the source and route of transmission
- I will request specific data to test those hypotheses
- You will provide feedback on the adequacy of the investigation

I am ready. Present an outbreak scenario.

[Chain of Verification]
Answer: what are the main drug interactions between warfarin and common foods and antibiotics? After answering, follow this verification protocol: 1. Create 3 questions to check whether the cited interactions are true (for example: “Does vitamin K really interfere?”). 2. Answer those questions independently. 3. Provide a revised and corrected final answer based on those checks.

[Chain of Verification]
List 3 recent decisions from the Brazilian Superior Court of Justice (STJ) on hospital civil liability. Follow the CoVe protocol: 1. Generate questions to confirm case number and year. 2. Answer whether the cases exist and actually concern the topic. 3. Deliver only the confirmed cases.

[Chain of Debate] You will act as two specialist physicians debating internally. Context: an elderly, frail patient with multiple comorbidities and a recent diagnosis of advanced cancer. Task: simulate a structured internal debate: Specialist A advocates aggressive treatment and Specialist B advocates early palliative care. Each must present their arguments, refute the other's and acknowledge limitations. Close with a balanced synthesis.

[Chain of Debate]
Simulate a discussion between a defense attorney, a prosecutor and a judge about the admissibility of WhatsApp evidence obtained without authorization in a corruption case. Debate “Fruit of the Poisonous Tree” versus “Public Interest.”

[Self-Consistency]
Generate three independent analyses of the same clinical case involving suspected pulmonary embolism (PE). Produce three separate diagnostic reasoning paths. Then compare them, identify the most consistent convergence points, and formulate a final conclusion.

[Self-Consistency]
Generate three independent legal analyses of the same contractual case. Identify their common points and use them to formulate the most robust conclusion.

[Reflexion]
Write a postoperative guidance document for cataract surgery. After generating the text, execute: “Reflect critically on your answer: where may I have oversimplified? Which assumptions may be wrong? What additional information would change my conclusion?” After that reflection, refine your original answer.

[Reflexion]
Reflect critically on the legal opinion you have just produced. Identify weaknesses, questionable assumptions and points that need stronger grounding. Revise the opinion after that analysis.

[Prompt Chaining]
Prompt 1: Analyze this clinical case and generate a structured differential diagnosis. Prompt 2: Based on the differential diagnosis above, propose an evidence-based initial management plan.

[Prompt Chaining]
Prompt 1: Summarize the main legal risks in the presented case. Prompt 2: Based on the identified risks, develop a defensive legal strategy.

[Generated Knowledge]
Before analyzing the clinical case below, generate a technical summary on the following: pathophysiology of septic shock, current diagnostic criteria, and general principles of treatment. Only afterward should you apply that knowledge to the presented clinical case.

[Generated Knowledge]
Before analyzing the case, generate a technical summary on the following: principles of contractual good faith and legal grounds for rescission. Then apply that knowledge to the concrete case.

[Least-to-Most]
First explain, in simple language, what heart failure is. Then: 1. Explain the pathophysiological mechanisms. 2. Differentiate HF with preserved versus reduced ejection fraction. 3. Relate that distinction to therapeutic choice.

[Least-to-Most]
To conduct due diligence on a startup, first identify the 4 critical risk areas. Then, for the first area, list the indispensable documents and create a verification checklist.

[Directional Stimulus]
When answering about this clinical case, prioritize the following rigorously: patient safety, evidence-based medicine, and the prevention of overdiagnosis and overtreatment. [Describe the clinical case here].

[Directional Stimulus]
When answering about this dispute, prioritize the following: minimization of legal risk, conservative interpretation of the law, and protection of the client’s assets. [Describe the case.]

[Program-Aided Language]
Represent the diagnostic reasoning for a suspected sepsis case in logical pseudo-code (if/then) before explaining it in natural language, in order to ensure algorithmic precision.

[Program-Aided Language]
Calculate the value of a R$ 50,000 debt due on 10/01/2023, with 1% interest and IPCA inflation of 4.5%. Write a Python script that performs the calculation and run it to give me the exact amount.

[ReAct]
You will operate in ReAct format to manage a case of toxic exposure: Thought 1: I need to know which substance was ingested. Action 1: Ask the user. Thought 2: Based on the response, I should check the antidote. Action 2: Provide the dose. Continue the cycle until stabilization.

And in a legal context:

[ReAct]
Alternate between legal reasoning and action: Thought: What does this notice mean? Action: Which document should be reviewed? Thought: What are the implications? Action: What response is recommended? Continue until a final strategy is reached.

If you have already noticed that "the same ChatGPT" behaves in very different ways depending on onde you use it? The logic is more or less the following: GPT is the model and it is placed in several different contexts: via the internet page, you access the chatGPT; on the computer, to generate code, you use codex — the model is the same, but the tools it has access to and the basic instructions that say what it can or cannot do are different. This environment around the model is what we call the harness.

The word comes from the harness (and the safety belt)

Harness, in English, is the harness of a horse, the chest harness of a dog, the climbing safety belt (and other examples, without prejudice). This structure worn around the person defines the context of their work, their safety, with which tools they will work... The figure is self-explanatory

In artificial intelligence, the harness is exactly that: the software layer that wraps the model and gives it tools, memory, permissions, validation and the capacity to act. It is the infrastructure that transforms "a text predictor" (the model) into "an assistant that does different things" (according to the harness).

The model alone does very little

Worth recalling what we saw in the section Basic mechanics: a language model (LLM), in isolation, does a single thing — it receives text and returns the most likely continuation, token by token. Alone, the model does not browse the internet, does not run code, does not open your files, does not remember yesterday's conversation and does not perform any action in the world. It only generates text.

Everything that seems like "the AI doing" — researching, reading a PDF, running a calculation, generating a spreadsheet, citing the exact source, remembering your preferences — is the harness working around the model.

How the harness works, in practice

When you submit a task, the harness executes more or less this cycle:

1. Receives the task and assembles the context (gathers your request, fixed instructions, relevant memory and, if applicable, excerpts from documents it went and fetched for you).
2. Sends it to the model and reads the answer.
3. If the model "asks for" a tool (search the web, run a piece of code, read a file, query an API), it is the harness that actually executes that action and returns the result to the model.
4. Repeats the cycle as many times as necessary — reason, act, observe the result, reason again — until the task is done.
5. Applies the brakes (guardrails): what is allowed, what requires your confirmation, what needs to be validated.
6. Delivers the result — often with a log of the steps and the sources (observability).

The engine and the vehicle analogy

It is the same logic as the bicycles and cars I used in the section AI models, now one floor up. Old-timer reference: with a Beetle engine, you can build both the Herbie and an ultralight aircraft. The model is the engine and the shell - what is around it - is the harness. In this bizarre example it is evident that the harness gives very different functions and characteristics to the model!

That is why "which is the best model?" is an incomplete question. In practice, you almost never use a pure model: you use a model + harness. When you compare ChatGPT with Gemini Notebook, or Claude on the website with Claude inside a code editor, much of the difference you feel is the harness, not the model.

From tool to agent

When the harness gives the model autonomy to plan, act in cycles and decide its own steps — and, sometimes, call on specialized "subagents" (one for code, one for images, one for research) —, that is what people mean when they say agente de IA or agentic AI. An agent, simply put, is model + harness with freedom to act.

Examples of harnesses you have probably already used

In short: the model is what thinks; the harness is what allows thinking with tools, acting with limits and being accountable for what was done. Without a harness, the model does not operate in the real world — and it is the harness that, in practice, you hire, configure and supervise.