Your First Custom Instructions

🔄 In Lesson 2, you learned how custom instructions sit at the system level and shape every conversation. Now let’s write your first set. By the end of this lesson, you’ll have working instructions installed on your platform of choice.

The RISEN Framework

RISEN gives you a structure so you’re not staring at a blank text field. Each letter stands for a component of good instructions:

• Role — Who is the AI? Define its expertise and perspective.
• Instruction — What should the AI do? Clear behavioral rules.
• Structure — How should output be formatted? Length, format, organization.
• Examples — Show, don’t tell. Provide 2-3 input-output pairs.
• Nuance — Edge cases, exceptions, constraints.

You don’t need all five for every instruction set. But the more components you include, the more consistent the AI’s behavior becomes.

Step-by-Step: Writing Your Instructions

Let’s build your instructions piece by piece. Grab a text editor — you’ll paste the result into your AI platform at the end.

R — Role

Start with who the AI should be. Not just a job title — include what matters about that role.

Vague (waste of characters):

“You are a helpful assistant.”

Specific (changes behavior):

“You are a senior marketing strategist who specializes in B2B SaaS. You think in terms of conversion funnels, customer segments, and data-backed decisions.”

The role isn’t decorative. Research from PromptHub shows that persona prompting acts as a “cognitive filter” — it primes the AI to retrieve patterns from relevant training data. A “senior developer” produces different code than a “beginner-friendly tutor.”

✅ Quick Check: You want the AI to help you write emails to enterprise clients. Which role instruction produces better results — “You are a helpful writer” or “You are a B2B sales copywriter who writes for CTO-level decision-makers at Fortune 500 companies”? (The second one. Specificity in the role primes the AI to use the vocabulary, tone, and structure appropriate for that audience.)

I — Instruction

Define behavioral rules. What should the AI always do? What should it never do?

Key principle: Tell the AI what TO DO, not what NOT to do. Anthropic’s official guidance: positive framing outperforms negative framing.

• Instead of: “Don’t use jargon” → “Use plain language a non-technical reader can understand”
• Instead of: “Don’t be verbose” → “Lead with the answer. Explain only if I ask follow-up questions.”

A starter set of behavioral rules:

• Lead with the answer, then explain if needed
• When I share work for feedback, be critical — point out problems, not praise
• If something is ambiguous, ask me one clarifying question instead of guessing
• Cite sources when making factual claims

S — Structure

Tell the AI how to format output. This is where most people see immediate improvement.

• Use bullet points for lists of 3+ items
• Keep responses under 300 words unless I ask for more
• Use headers (##) to organize responses longer than 200 words
• For code: brief explanation first, then a single code block

E — Examples

This is the power move. Two to three examples teach the AI more than a paragraph of rules.

When I ask for feedback on writing:

• Don’t say: “This is great! Maybe consider adding…”
• Do say: “Paragraph 2 is weak. The claim about market size needs a source, and the transition from feature to benefit is missing.”

Examples make abstract instructions concrete. The AI sees the pattern and replicates it.

N — Nuance

Handle edge cases and exceptions. This prevents your instructions from being too rigid.

• If I write in a language other than English, respond in that language
• For creative tasks (brainstorming, fiction), relax the structure rules — be more free-flowing
• When I say “quick,” limit the response to 2-3 sentences max

✅ Quick Check: You’ve written instructions saying “always use bullet points.” But sometimes you want prose — like for cover letters or blog posts. How do you fix this? (Add a Nuance rule: “Use bullet points by default. Switch to prose for tasks that require flowing text — emails, letters, essays, blog posts.”)

Putting It Together: Your First Instructions

Here’s a complete example for a marketing professional. Adapt it to your role:

ROLE: You are a senior B2B marketing strategist. You think in
conversion funnels, customer segments, and ROI.

RULES:
- Lead with the answer, then explain
- Be direct and critical when reviewing my work
- Ask one clarifying question if my request is ambiguous
- Default to American English

FORMAT:
- Bullet points for lists, prose for analysis
- Responses under 300 words unless I request more
- Use ## headers for anything over 200 words

EXAMPLES:
When I ask "is this headline good?":
- Not this: "It's good! Maybe try..."
- This: "Weak. 'Revolutionize' is a cliché. Try: '[Specific benefit] in [timeframe]'"

NUANCE:
- For brainstorming: skip structure, go rapid-fire
- For data analysis: include methodology notes
- When I say "quick" → 2-3 sentences max

Exercise: Install Your Instructions

1. Copy the template above (or write your own using RISEN)
2. Customize it for your role, preferences, and common tasks
3. Install it:
   • ChatGPT: Settings → Personalization → Custom Instructions
   • Claude: Create a Project → Set Project Instructions
   • Gemini: Create a Gem with your instructions
4. Test with three different prompts. Notice the difference.

Key Takeaways

• RISEN (Role, Instruction, Structure, Examples, Nuance) gives you a framework to fill the blank text field
• Specific roles produce better output than generic ones — “senior Python developer” beats “helpful assistant”
• Positive framing (“use plain language”) outperforms negative framing (“don’t use jargon”)
• Two to three examples teach the AI more effectively than paragraphs of rules
• Add flexibility with Nuance rules — “by default… unless” prevents rigid behavior

Up Next

Your instructions work. But you’ve only scratched the surface. In Lesson 4, we’ll cover advanced techniques: few-shot chaining, conditional logic, multi-persona instructions, and how to use XML tags to organize complex instruction sets.