System Prompt Design: The Art of Good Instructions

L3 Lesson 5 of 5 — Context as Infrastructure

The Bridge Between Knowledge and Impact

You now know three ways to build persistent context: Claude Projects, Custom GPTs, M365 Copilot. All three share one thing: Quality rises and falls with the quality of your instructions. Whether you write custom instructions in a Claude Project, configure a Custom GPT, or craft a complex prompt in a single chat — the principles are the same.

This lesson is the synthesis of L3: How to build a System Prompt that works.

Anatomy of a Strong System Prompt

The golden rule comes from Anthropic:

“Show your prompt to a colleague with minimal context and ask them to follow it. If they’d be confused, the AI will be too.”

An AI model is like a brilliant new hire: Highly capable, but without context on your norms, style, and expectations. The clearer you explain what you want, the better the result.

The Seven Building Blocks

Not every prompt needs all seven. But knowing them lets you consciously decide which to use:

Building Block	Purpose	Example
Role	Focuses behavior and expertise	”You are a senior strategy consultant.”
Context	Explains the why behind the rules	”The response will be read aloud — no ellipses.”
Task	What specifically should be done	”Analyze quarterly reports and identify trends.”
Format	How the output should look	”3–5 bullet points, max 2 sentences each.”
Constraints	Boundaries and limitations	”Only use information from the provided documents.”
Examples	Concrete input/output pairs	One good example shows more than ten rules.
Tone	Language style and register	”Professional but approachable, use ‘you.’”

The Most Important Insight

Context explains the why. And the why is the most powerful building block.

Compare:

“Don’t use ellipses.” → The AI follows the rule, but only literally.
“The response will be read by a text-to-speech engine. Don’t use ellipses because the engine can’t pronounce them.” → The AI understands the reason and also avoids other TTS problems you didn’t explicitly mention.

Three Examples: From Draft to Strong Prompt

Example 1: Technical Editor

You are an experienced technical editor. Your audience has solid
foundational knowledge but no deep specialization.

Style:
- Professional but approachable
- Prefer active voice
- Explain technical terms on first use
- Plain English with industry-standard terminology

Format:
- Flowing prose with clear paragraph structure
- Subheadings only for texts over 500 words

Quality:
- Support claims with source references
- When uncertain: "I'm not confident here, but..."
- No marketing language or superlatives without evidence

Why it works: Clear role with audience definition. Positively stated style rules. Honesty constraint for uncertainty.

Example 2: Data Analysis Sparring Partner

You are an analytical sparring partner for business data.

Workflow for every analysis:
1. Summary: What do the data show? (2-3 sentences)
2. Key findings: The 3 most important patterns
3. Context: What might explain the data?
4. Recommended action: What would I do?
5. Open questions: What data are missing?

Rules:
- Never present correlation as causation
- Always include absolute numbers alongside percentages
- No recommendations without data-backed reasoning

My context:
Product Manager, B2B SaaS, 200 customers.
Key metrics: MRR, Churn Rate, NPS, Feature Adoption.

Why it works: Structured workflow with numbered steps. Hard analytical rules. Personal context makes recommendations relevant.

Example 3: Customer Support Assistant

You help me respond to customer inquiries professionally.
Our product: project management software for teams.

Tone:
- Friendly and solution-oriented
- Direct — no filler phrases like "I'd be happy to help you"
- Formal register unless the customer uses informal first

Process:
1. Restate the problem in your own words
2. Suggest a solution or next step
3. Ask if that helps or offer further assistance

Escalation:
If you're unsure or the issue is technical:
- Be honest: "I need to check this internally."
- Give a concrete timeframe
- DO NOT invent solutions for technical problems you don't understand

Why it works: Specific tone guidance (“no filler phrases like…”). Escalation rules prevent hallucination.

The Five Most Common Anti-Patterns

1. Vague Instructions

Problem	Better
”Be helpful"	"Answer in 2–3 sentences with one example."
"Format nicely"	"Flowing prose, bullet points only for 4+ items."
"Be professional"	"Use technical terms where needed, define abbreviations on first use.”

2. Contradictory Rules

“Keep it short.” + “Explain every point in detail.” → The AI guesses what you mean.

Solution: Prioritize. “Default to concise (2–3 sentences). For technical explanations, go deeper with an example.”

3. Nothing but Prohibitions

“Do NOT use Markdown.” — “No long sentences.” — “Do NOT hallucinate.”

Better: Say what you want, not what you don’t. “Respond in flowing prose with short sentences.”

4. Overloaded Prompts

System prompts with 500+ lines lead to declining compliance, contradictions, and high token costs. Stay under 200 lines. For more complex setups: break into modular blocks.

5. Rules without Reasons

“NEVER use ellipses.” → The AI follows the rule but doesn’t understand why.

“The response will be read by a TTS engine. No ellipses because the engine can’t pronounce them.” → The AI generalizes and avoids similar problems automatically.

L3 Summary

Build persistent context for recurring tasks — explain your job once
Choose the right tool for the context: workspace, specialized tool, or embedded AI
Write system prompts positively: what the AI should do, not what it shouldn't
Explain the why behind rules — the AI generalizes better
Iterate: start minimal, observe, correct with intention

Start every chat from zero when context and task repeat
Cram more than 200 lines into a system prompt — less with structure beats more without
Upload confidential data to insecure knowledge stores (e.g., Custom GPT Knowledge Files)
Try to make everything perfect at once — the compound effect takes time
Force one tool for everything — different tasks need different tools

Your L3 Checklist

Before moving to L4, you should be able to answer these with “yes”:

I understand the difference between transactional and persistent AI usage
I’ve set up at least one persistent workspace (Claude Project, Custom GPT, or equivalent)
I can name the strengths and weaknesses of Claude Projects, Custom GPTs, and M365 Copilot
I can write a system prompt with role, context, rules, and an example
I know the five anti-patterns and how to avoid them

What’s Next?

In L4 — AI as Coworker, you take the next step: From context to delegation. You’ll learn how to not just formulate tasks better, but hand over entire workflows to AI — with the right balance of trust and control.

The paradigm shift of L4: From “AI answers me” to “AI works for me.”