🎯 Why Prompting Matters
A bad prompt gets a bad answer. A good prompt gets a great answer. The difference? A 2-minute rewrite.
This is the highest ROI skill you can learn. Get prompting right, and Claude becomes 10x more useful. Everything else in Phase 3 and 4 depends on this.
🏗️ The 4-Part Prompt Structure
Every good prompt has these 4 parts. Use them in order:
ROLE:
"You are a [specific expertise]. You know [specific domain knowledge]."
TASK:
"Write/Build/Debug/Analyze [specific output]. The output should [specific requirements]."
CONTEXT:
"Here's the background: [relevant information]. Constraints: [limits]."
FORMAT:
"Return the output as [format]. Structure it like [example or structure]."
❌ Bad Prompt
Write me a Python script.
✅ Good Prompt
You are an expert Python developer. Write a Python script that parses CSV files and exports them as JSON. It should handle errors gracefully and log all operations. Return the complete code with comments.
Pro Tips
- Be specific. "Make it better" → "Add error handling for network timeouts"
- Show examples. "Format like this: ..." → Claude copies your style
- Add constraints. "Keep it under 200 tokens" → Claude optimizes
- State the WHY. "This is for a startup MVP" → Claude prioritizes speed over perfection
📈 5 Levels of Prompting
Master these one by one. Level 1 → Level 5 = 100x better results.
Level 1: Basic
Single sentence. No context. "Write a poem." Works but messy output.
Level 2: Structured
Use the 4-part framework. Claude gives you what you asked for.
Level 3: Examples
Show Claude examples of good output. It copies your style exactly.
Level 4: Reasoning
Ask Claude to "think out loud" first. Better for complex problems.
Level 5: Expert
Combine all of the above. Use CLAUDE.md for context. Perfect results every time.
💡 Level 3: Show Examples (The Game Changer)
Claude doesn't just follow instructions—it copies patterns. Show it what you want, and it replicates it perfectly.
You are a technical writer. Rewrite this in YOUR style:
EXAMPLE OF YOUR STYLE:
- Title: starts with emoji + action verb
- Tone: conversational, not corporate
- Length: 2-3 sentences max
- Format: bullets with arrows
Here's the text to rewrite: [text here]
Match my style exactly. Use emoji, short sentences, action verbs.
Why this works: Claude learns from examples better than from instructions. Show, don't tell.
🧠 Level 4: Chain of Thought (The Reasoning Hack)
For complex problems, ask Claude to think before answering.
I need to debug this Python error: [error message]
Please:
1. Think out loud about what could cause this
2. Consider the most likely causes first
3. Propose 3 solutions from most to least likely
4. Show code for each solution
Think through this step-by-step before answering.
Why this works: "Think first" mode takes more tokens but gives way better answers for complex problems.
🎯 The CLAUDE.md System (Level 5)
This is the secret that changes everything.
CLAUDE.md is a file you create in your project. It stores rules, context, and preferences that Claude remembers across all your messages. No more repeating yourself.
What goes in CLAUDE.md:
- Code style preferences: "Use TypeScript, no semicolons, 2-space indent"
- Project context: "This is a React app for [purpose]. Stack: [technologies]"
- Rules: "Never use X library. Always add error handling. Include tests."
- Target audience: "This is for beginners. Explain everything."
- Quality standards: "Add 100% test coverage. Performance critical—optimize first."
Example CLAUDE.md:
# Miru Project
## What This Is
- WhatsApp AI assistant for UK daily life
- Python + Flask backend, Supabase DB, Twilio integration
- Deployed to Railway (auto-deploy on git push main)
## Rules (ALWAYS FOLLOW)
- Use Python 3.11+
- Add error handling for all API calls
- Never hardcode secrets (use .env)
- Test with real phone numbers, not fake ones
- Always include docstrings
## Code Style
- 2-space indent
- Type hints everywhere
- No wildcard imports
- Class names: PascalCase, functions: snake_case
## Team
- Vikram built this. Know his style: pragmatic, ship fast, focus on user value
- When unsure about design, ask: "Would Vikram do this?"
## Next Priority
- Fix school comms polling
- Add motivation features
- Deploy by Friday
How to use it: Just put CLAUDE.md in your project root. Claude Code reads it automatically and follows all your rules.
⚠️ Common Mistakes (Don't Do These)
❌ Too Vague
"Make this function faster"
✅ Specific
"This function processes 10k records and takes 30s. Target: under 5s. Profile shows 80% time in database queries."
❌ Rambling
"I have this problem where the thing isn't doing the other thing and I'm not sure why but it might be because of..."
✅ Concise
"Error: 'NoneType' object is not iterable. Line 42. Variable 'x' is None when it shouldn't be."
❌ No Context
"Debug this" [pastes code snippet]
✅ With Context
"This code runs on Railway with Supabase. When user uploads a file over 50MB, it fails. I've added logging—see attached."
The "Prompt Audit" Checklist
- Is the ROLE clear? (Who should Claude be?)
- Is the TASK specific? (What exactly do you want?)
- Is there CONTEXT? (What's the background?)
- Is the FORMAT defined? (How should it be returned?)
- Can you read it to someone else in 30 seconds?
🎯 Practice Exercise
Rewrite this bad prompt into a good one:
Write me a function to do stuff with data.
Good answer should have:
- ✓ ROLE: "You are a [type] developer"
- ✓ TASK: "Write a function that [specific behavior]"
- ✓ CONTEXT: "It takes [input], returns [output]"
- ✓ FORMAT: "Return [language] code with [comments/tests]"
💪 Key Takeaway
Perfect prompt = 10 seconds thinking + 2 minutes writing.
The time you spend writing a good prompt saves you 10x the time dealing with bad output. It's the highest-ROI skill you can learn this week.