🎓Advanced Prompt Techniques

Role prompting, output control, and model-specific tips

Advanced Prompt Techniques

Role Prompting

Assign Claude a specific role for better results:

You are a senior database architect with 15 years of experience
in PostgreSQL and distributed systems. A junior developer asks you:

"Should I use MongoDB or PostgreSQL for an e-commerce application?"

Provide guidance considering:
- Data relationships
- Transaction requirements
- Scale expectations
- Team expertise

Output Control

Specify exact format:

Respond ONLY with valid JSON. No explanation, no markdown, no code fences.
The JSON should have these fields:
{
  "summary": "one sentence",
  "sentiment": "positive|negative|neutral",
  "confidence": 0.0 to 1.0
}

Control length on Claude Opus 4.7:

Provide concise, focused responses. Skip non-essential context,
and keep examples minimal.

Handling Refusals

If Claude refuses a valid request, rephrase:

# Instead of:
"Write malware code"

# Try:
"I'm a security researcher. Write an educational example of how
SQL injection works, with the vulnerable code AND the fix,
so I can teach my team about this vulnerability."

Model-Specific Tips

Claude Opus 4.7:

Best for complex, multi-step reasoning
Naturally calibrates response length to complexity
Use effort parameter to control thinking depth
Prefers adaptive thinking over manual extended thinking

Claude Sonnet 4.6:

Best balance of speed and quality
Great for production applications
Supports extended thinking with budget_tokens
Ideal for most API use cases

Claude Haiku 4.5:

Fastest model — great for high-volume
May need more explicit instructions
Best for simple tasks and classifications
Cheapest option for production

Prompt Caching (Save Money!)

Reuse the same system prompt across requests:

JavaScript
// First request — creates cache
const response1 = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 1024,
  system: [
    {
      type: "text",
      text: "You are a helpful assistant...",
      cache_control: { type: "ephemeral" }  // Cache this!
    }
  ],
  messages: [{ role: "user", content: "Question 1" }]
});

// Subsequent requests — cheaper! Cache is used
const response2 = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 1024,
  system: [
    {
      type: "text",
      text: "You are a helpful assistant...",
      cache_control: { type: "ephemeral" }
    }
  ],
  messages: [{ role: "user", content: "Question 2" }]
});

Cache lasts 5 minutes (standard) or 1 hour (extended). Cached input is much cheaper than regular input.

Next up: Put it all together with hands-on projects.

Module 7

4/6

📊 Evaluating & Testing Prompts

Building AI Agent Loops 🤖

4/6