PTENES
MODULE 2.1

🎯 Creating your custom assistant

Build an AI assistant that understands your profession, your audience, and the way you work. Leave this module with a system prompt ready to use.

6
Topics
40
Minutes
Practical
Level
Hands-on
Type
1

🎯 What a custom AI assistant is

One custom AI assistant is an artificial intelligence configured with specific instructions about who you are, what you do, and how you work. Technically, this is called system prompt (system prompt): a block of instructions the AI reads before each conversation that defines its behavior.

The difference between using ChatGPT “out of the box” and a customized assistant is like the difference between hiring an intern on their first day and having a colleague who’s worked with you for 2 years. The intern needs everything explained. Your colleague already knows your jargon, audience, and style.

Generic Assistant "I can help with a variety of tasks..." Generic responses Default tone Needs an explanation every time Custom Assistant "Dr. Silva, about the patient’s case..." Responses with professional context Tone tailored to your field It already knows who you are

✓ A custom assistant DOES

  • ✓Automatically uses terminology from your field
  • ✓Maintain the tone you defined (formal, technical, accessible)
  • ✓Follows the ethical rules of your profession
  • ✓Produce it in the format you need (report, opinion, email)

✗ A generic assistant DOES NOT

  • ✗Doesn't know your specialty or audience
  • ✗Replies in a generic tone that doesn't sound like you
  • ✗Ignore regulatory restrictions in your field
  • ✗Produces a standard format that you need to rework
70%

less rework

3x

faster

Tone

consistent

1x

set it up, use it regularly

2

🏗️ Anatomy of the system prompt

O system prompt (system prompt) is the instruction text the AI reads before each conversation. Think of it as your assistant's "field manual." It has 5 parts that, together, define exactly how the assistant behaves.

1. Identity Who you are 2. Audience Who you speak to 3. Rules What You Can and Can’t Do 4. Format How to respond 5. Knowledge Your database Your Assistant customized and ready
1

Identity

Who the assistant is and what its specialty is

Example: "You are an assistant specializing in civil law, with a focus on business contracts. Your name is JurisBot."

2

Target audience

Who the assistant will create content for

Example: "Your clients are small and medium-sized business owners without legal training. Use clear language and avoid unnecessary Latin terms."

3

Behavior rules

Ethical and operational limits

Example: "Never make a medical diagnosis. Always indicate that the opinion must be validated by the professional. Never invent case law."

4

Response format

How to structure the outputs

Example: "Always answer in bullet points. Use no more than 3 paragraphs. When drafting legal documents, use the template [HEADER > LEGAL BASIS > REQUEST]."

5

Knowledge base

Documents and references the assistant should use

Example: "Use the Civil Code (2002), STJ summaries on contracts, and the attached contract templates as references."

3

👩‍⚕️ Practical example: assistant for a doctor

Let's build a complete assistant for Dr. Mariana Silva, a general practitioner focused on preventive medicine. The prompt below is real and functional. Copy, adapt, and paste it into ChatGPT or Claude.

Complete system prompt

You are the assistant to Dr. Mariana Silva, a primary care physician with 15 years of experience in preventive medicine. AUDIENCE: Adult patients ages 30-60, middle class, seeking prevention and quality of life. Use clear, welcoming language without unnecessary medical jargon. RULES: - NEVER diagnose. Say "consult your doctor for confirmation" - NEVER recommend specific medications - Always suggest professional follow-up for concerning symptoms - Use an empathetic but direct tone FORMAT: - Use bullet points when possible - Maximum 4 paragraphs per response - Include "Suggested next step:" at the end KNOWLEDGE: 2024 Ministry of Health guidelines, INCA screening protocols, SUS preventive procedures table.

💡 Copy and use

Paste this prompt into the “Custom Instructions” field in ChatGPT or the “System Prompt” in Claude. Then test it with: “Female patient, 45, wants to know which preventive screenings to get this year.”

📊 Results: Before vs. After

  • Without a prompt: "Mammogram, Pap test, complete blood count... recommended" (cold, technical list)
  • With prompt: "Hi! At 45, it’s a good idea to stay up to date on 3 important preventive screenings..." (warm tone, audience-appropriate language)
4

👩‍⚖️ Practical example: assistant for a lawyer

Now let's build the assistant for Dr. Carla Mendes, a civil lawyer specializing in contracts and business law. Compare it with the medical example and note what changes and what stays the same in the structure.

Complete system prompt

You are the legal assistant to Dr. Carla Mendes, a civil attorney with 12 years of experience in business contracts and corporate law. AUDIENCE: Small and medium-sized business owners who need legal guidance in accessible language. Avoid Latin terms; explain technical terms. RULES: - NEVER make up case law or statutory provisions - Always cite the legal source when mentioning legislation - Include "This is a draft — have your attorney review it before use" in any document - Respect attorney-client confidentiality: do not share data between clients - Follow the OAB rules on advertising FORMAT: - Legal documents: HEADER > FACTS > LEGAL BASIS > REQUEST - Emails: formal but accessible, maximum 3 paragraphs - Legal opinions: SITUATION > ANALYSIS > CONCLUSION > RECOMMENDATION KNOWLEDGE: Civil Code (2002), CDC, Corporations Act, CLT (labor clauses in contracts), most cited STJ summaries in contract law.

✓ What is UNIVERSAL (every profession)

  • ✓Identity block with name and specialty
  • ✓Clearly define the target audience
  • ✓Negative rules ("never do X")
  • ✓Structured output format

✎ What CHANGES (specific to the field)

  • ✎Technical terminology (medical vs. legal)
  • ✎Ethical restrictions (diagnosis vs. case law)
  • ✎Document templates (expert report vs. petition)
  • ✎Knowledge base (protocols vs. legislation)

💡 Tip: cross-test

After creating your prompt, ask the same question to the generic assistant and the customized one. The difference is the best argument for why the effort was worthwhile. Test prompt: "A client wants to terminate a service agreement. What should I check before advising them?"

5

🧪 Build yours now

Here is the universal template. Works for any independent professional. Copy it, fill in the fields in brackets, and paste it into your AI tool.

Universal system prompt template

You are the assistant to [YOUR NAME], a [YOUR PROFESSION] with [X] years of experience in [YOUR SPECIALTY]. AUDIENCE: [DESCRIBE YOUR CLIENTS/PATIENTS - age, profile, level of technical knowledge]. Use [FORMAL/ACCESSIBLE/TECHNICAL] language, [TONE INSTRUCTIONS]. RULES: - NEVER [ETHICAL RESTRICTION 1 - e.g., diagnose, make up data] - NEVER [ETHICAL RESTRICTION 2] - Always [POSITIVE RULE 1 - e.g., cite sources, suggest validation] - Always [POSITIVE RULE 2] FORMAT: - [DOCUMENT TYPE 1]: [STRUCTURE - e.g., HEADER > BODY > CONCLUSION] - [DOCUMENT TYPE 2]: [STRUCTURE] - General responses: [FORMATTING RULE - e.g., maximum 3 paragraphs, use bullet points] KNOWLEDGE: [YOUR REFERENCES - standards, laws, protocols, guides the assistant should follow].

🎯 Objective

Fill in the template above with your information and paste it into ChatGPT (Custom Instructions) or Claude (System Prompt).

How to check: After pasting it, ask a question typical of your day-to-day work. If the answer already has the tone and format you expect, it’s working. If not, revise the RULES and FORMAT sections.

Where to Paste the Prompt (by Tool)

1

ChatGPT

Settings > Personalization > Custom Instructions > "How would you like ChatGPT to respond?"

2

Claude

Projects > New Project > Instructions (field at the top) or paste it at the start of the conversation

3

Gemini

Google AI Studio > System Instruction (field at the top of the chat)

6

✅ Test and refine

Your assistant won’t be perfect on the first try. The secret is test with real questions and keep adjusting the system prompt. Here's a checklist of 5 tests that reveal whether the assistant is calibrated.

5-test checklist

1.
Tone test: Ask for a reply to a client. Does the tone sound like you?
2.
Jargon test: Ask for a technical document. Does it use the terms from your field correctly?
3.
Limit test: Ask for something you prohibited (e.g., a diagnosis). Does it refuse?
4.
Format test: Ask for a document. Does it follow the structure you defined?
5.
Audience test: Ask for an explanation. Is the language appropriate for your audience?

Prompt to try (copy and use)

I'll run 5 quick tests to calibrate you. Answer each one: 1. A client/patient asks about [COMMON SITUATION IN YOUR FIELD]. Respond as you would. 2. Draft a [TYPICAL DOCUMENT IN YOUR FIELD] about [SIMPLE TOPIC]. 3. [QUESTION THAT SHOULD BE REFUSED - e.g., "What medication should I take?"] 4. Summarize the following case in 3 bullet points: [DESCRIBE A SIMPLIFIED REAL CASE] 5. Explain [TECHNICAL CONCEPT IN YOUR FIELD] to a layperson.

How to check: If 4 out of 5 responses have the expected tone and format, your prompt is good. If fewer than 3 do, revise the RULES and FORMAT sections.

⚠️ Common mistakes in your first prompt

  • ✗Prompt that's too vague: “Be a good assistant” (doesn't work)
  • ✗Prompt that's too long: more than 500 words dilutes the AI's attention
  • ✗Without negative rules: AI needs to know what NOT to do
  • ✗Without a defined format: AI chooses a different format for each response

📋 Module Summary

✓
Custom vs. generic assistant - the difference between constant rework and real productivity
✓
5 parts of the system prompt - Identity, Audience, Rules, Format, Knowledge
✓
Real examples - doctor and lawyer with functional prompts
✓
Universal template - copy, fill in, and use in any tool
✓
Testing checklist - 5 questions to check whether it's working

Next Module:

2.2 - Teaching AI the context of your profession

← Back to Track Next Module →