Writing Instructions That Actually Work

Instructions are the heart of every custom GPT. A GPT with weak instructions is just ChatGPT with a name. A GPT with strong instructions is a reliable, consistent tool that impresses every user.

This is the most important lesson in the course.

🔄 Quick Recall: In the previous lesson, you explored the GPT Builder interface and built a simple GPT. Now you’ll learn to write instructions that transform a basic GPT into a professional-grade tool.

The ROLE-RULES-FORMAT Framework

Great GPT instructions follow a three-part structure:

ROLE — Who is the GPT? What’s its expertise and personality? RULES — What should it do and not do? What are the boundaries? FORMAT — How should responses look? What’s the output structure?

Here’s a simple example for a meeting notes GPT:

ROLE:
You are a professional meeting notes assistant. You specialize in
turning messy meeting transcripts into clear, actionable summaries.

RULES:
- Always extract action items with assigned owners and deadlines
- Separate decisions made from topics only discussed
- If no owner is specified for an action item, flag it as "UNASSIGNED"
- Never add information that wasn't in the original transcript
- Keep summaries under 500 words unless the user requests more detail

FORMAT:
Structure every response as:
## Meeting Summary
[2-3 sentence overview]
## Key Decisions
[Bulleted list]
## Action Items
[Table: Action | Owner | Deadline]
## Discussion Notes
[Brief notes on topics discussed but not decided]

This GPT will produce consistent, well-structured meeting notes every single time — because the instructions leave nothing to interpretation.

✅ Quick Check: What are the three parts of the ROLE-RULES-FORMAT framework? (Answer: Role defines who the GPT is, Rules define what it should and shouldn’t do, and Format defines how responses should look.)

Writing the Role Section

The role section sets the GPT’s identity and expertise. Be specific:

The more specific the role, the more focused and useful the responses. A GPT that tries to be everything helps with nothing.

Writing Effective Rules

Rules come in two types: positive (what the GPT should do) and negative (what it shouldn’t).

Positive rules:

• “Always ask for the target audience before writing”
• “Include at least 2 examples with every explanation”
• “Start every response with a one-sentence summary”

Negative rules (prevent common failure modes):

• “Don’t apologize or say ‘I’m sorry’ — provide solutions instead”
• “Don’t ask more than 2 clarifying questions before giving an initial response”
• “Never make up statistics or cite sources you haven’t verified”
• “Don’t repeat the user’s question back to them”

Negative rules are especially important because GPTs have default behaviors that may not match your goals: being overly apologetic, asking too many questions, or producing verbose responses.

✅ Quick Check: Why are negative instructions (“Don’t do X”) important in GPT instructions? (Answer: GPTs have default behaviors like apologizing, being verbose, or asking too many questions. Negative instructions prevent these unwanted behaviors.)

Trigger-Instruction Pairs

For GPTs that handle multiple types of requests, use trigger-instruction pairs — “When the user does X, respond with Y”:

TRIGGERS:

When the user uploads a resume:
→ Review it for ATS compatibility, formatting, and impact language.
→ Provide a score out of 100 and 5 specific improvements.

When the user provides a job description:
→ Analyze the key requirements and suggest resume modifications.
→ List the top 5 keywords to include.

When the user asks for a cover letter:
→ Ask which job description to target.
→ Write a 3-paragraph cover letter matching the role.

When the user asks something unrelated to job applications:
→ Politely redirect: "I specialize in resume and cover letter help.
   How can I help with your job application?"

This structure makes the GPT predictable. Users get the right response regardless of how they phrase their request.

Formatting with Markdown

Use Markdown in your instructions — GPTs read it better than plain text:

• Headers (##) to separate sections
• Bullet points for lists of rules
• Bold for emphasis on critical instructions
• Code blocks for exact output templates
• Tables for structured output formats

OpenAI’s own documentation confirms that “incorporating Markdown syntax and structured formatting can significantly improve the readability and precision of your prompts.”

Common Instruction Mistakes

Practice Exercise

Write instructions for a GPT of your choice using the ROLE-RULES-FORMAT framework:

1. Define the role in 2-3 sentences (be specific about expertise and personality)
2. Write 5 positive rules and 3 negative rules
3. Define the output format for the most common response type
4. Add 2 trigger-instruction pairs for different user scenarios
5. Paste into the Configure tab and test with 5 different prompts

Key Takeaways

• Use the ROLE-RULES-FORMAT framework for every GPT’s instructions
• Be specific in the role section — narrow expertise outperforms broad claims
• Include both positive and negative rules to prevent unwanted behaviors
• Use trigger-instruction pairs for GPTs that handle multiple request types
• Format instructions with Markdown for better AI parsing

Up Next

In the next lesson, you’ll learn to upload knowledge files that give your GPT domain expertise — turning it from a generic assistant into a specialist that knows your data.