🧩 Tokens, context, and window
AI doesn’t read words the way we do. It splits text into chunks called tokens (a token is approximately 3/4 of a word in Portuguese). The phrase "Good morning, how can I help?" has about 8 tokens. This matters because AI has a limit on how many tokens it can process at once.
A context window and the AI's “working memory.” Everything you write AND everything it replies needs to fit in this window. When the conversation gets too long, the AI starts to “forget” what was said at the beginning. That's why new conversations about different topics work better than one endless conversation.
💡 In practice
If you're using AI to review a long contract and then ask it to write an email about something else in the same conversation, the contract will "fall out" of the context window. Start a new conversation for each different task.
🎭 Persona and tone
A persona and the “role” you assign to AI. When you say “you are a tax attorney with 20 years of experience,” the AI adjusts its vocabulary, level of detail, and tone to match that persona. The more specific the persona, the better suited the result.
✓ With a defined persona
"You are a cardiologist explaining test results to a layperson."
Result: accessible language, a welcoming tone, everyday analogies.
✗ Without a persona
"Explain the results of this test."
Result: technical language and a generic tone may confuse the patient.
Prompt to try now
You are a [YOUR PROFESSION] with [X] years of experience.
Your tone is [formal/warm/direct/technical].
Explain [CONCEPT IN YOUR FIELD] to a client with no specialized knowledge.
Use no more than 100 words and avoid jargon.
📝 Prompt and Instruction
Prompt is the technical term for the instruction you give AI. It can be a short sentence ("summarize this document") or a detailed paragraph with context, the desired format, and constraints. The golden rule: the clearer and more specific the prompt, the better the result.
Too vague. The AI doesn't know who it's for, what it's about, or what tone to use.
Better, but it needs tone, length, and context details.
Be specific: recipient, context, reason, action, format.
🔄 Iteration — refining is part of the process
Iteration is the process of refining AI’s output by asking for adjustments. The first response is rarely perfect, and that’s normal. Professionals who iterate get significantly better results than those who accept the first version.
🔄 Iteration phrases that work
First attempt
"Write a business proposal for my firm."
Result: generic, impersonal, with no firm data.
After 3 iterations
"More formal. Add our 24-hour service differentiator. Replace the numbers with actual data from the last quarter."
Result: personalized, persuasive, ready to send.
💡 The right mindset
Think of AI as an automatic first draft. You wouldn't accept an intern's first draft without reviewing it. It's the same with AI: the value is in refinement cycle, not in the first response.
⚖️ Hallucination — when AI makes things up
Hallucination is the technical term for when AI generates false information with complete confidence. It may cite a law that doesn’t exist, invent case law, fabricate research data, or create a fake bibliographic reference. All with the same fluency as genuine text.
⚠️ Real examples of hallucinations
- ✗Lawyers in the U.S. cited fake case law generated by ChatGPT in an actual court filing (Mata v. Avianca case, 2023)
- ✗AI made up numbers for Brazilian law articles that don't exist
- ✗Fabrication of scientific studies with fictional authors, journals, and DOIs
- ✗Invented statistical data with seemingly credible sources
✓ How to protect yourself
- ✓Always verify citations, laws, and data against the primary source
- ✓Ask AI to provide the sources (and check that they exist)
- ✓Use AI for drafts, never as a final version without review
- ✓Be wary of data that is overly precise or too convenient
✗ Signs of hallucination
- ✗Law article numbers in an unusual format
- ✗Citations from authors you’ve never heard of
- ✗Statistical data with no verifiable source
- ✗Information that seems "too good to be true"
🛡️ Privacy and data
As an independent professional, you deal with confidential information of clients every day. Before using any AI, you need to understand what happens to the data you enter. The good news: there are safe ways to use AI without compromising professional confidentiality.
🛡️ Golden rules for privacy
- •Always anonymize: replace real names with fictional ones ("Client A," "Dr. Silva")
- •Never paste CPF, RG, or banking information or information that identifies a person
- •Use paid plans (ChatGPT Plus, Claude Pro): they don't use your data to train the model
- •Disable history if you're using the free version (option in settings)
- •When in doubt, don't share. Ask AI to work with a similar hypothetical case
Safe prompt for reviewing documents
Review the contract below. Names and amounts have been anonymized.
CLIENT: Company Alpha
CONTRACTOR: Company Beta
VALUE: R$ XXX.XXX,XX
[paste the anonymized text here]
Identify clauses that could pose a risk to the Contracting Party.
💡 LGPD and AI
The General Data Protection Law (LGPD) applies to the use of AI. You are the controller of the data you share with the platform. The AI provider is the operator. Responsibility for the proper use of your clients' data is your.
✓ Module Summary
Next Module:
1.3 - Choosing your tool — ChatGPT, Claude, Gemini