[ ABORT TO HUD ]
SEQ. 1
SEQ. 2
SEQ. 3
Few-Shot & Chain-of-Thought
Teaching Agents by Example
Two of the most powerful prompting techniques: Few-Shot Prompting (showing examples) and Chain-of-Thought (demonstrating reasoning steps).
Few-Shot Prompting for Tool Use
## Tool Usage Examples
User: "What's the weather in London?"
Thinking: User wants weather data. I should use get_weather.
Action: get_weather({"location": "London, UK"})
Result: {"temp": 12, "condition": "cloudy"}
Response: "It's 12°C and cloudy in London."
User: "Tell me a joke"
Thinking: General request, no tool needed.
Response: "Why do programmers prefer dark mode?..."
Chain-of-Thought Comparison
| Technique | When to Use | Token Cost | Quality Boost |
|---|---|---|---|
| Zero-Shot CoT | "Think step by step" | Low (+50 tokens) | +20-30% |
| Few-Shot CoT | Provide reasoning examples | Medium (+200) | +40-60% |
| Structured CoT | Force specific format | Medium (+300) | +50-80% |
| Extended Thinking | Claude native feature | Separate budget | Highest |
Structured CoT Template
Before answering, reason through these steps: 1. **Understand:** What is the user asking for? 2. **Gather:** What information do I need? 3. **Plan:** What's my step-by-step approach? 4. **Execute:** Carry out the plan. 5. **Verify:** Does my answer address the question?
🎯 Pro Tip: Always include a NEGATIVE example in few-shot prompts - showing when the agent should NOT use a tool. This dramatically reduces unnecessary tool calls.
SYNAPSE VERIFICATION
QUERY 1 // 2
Why should few-shot prompts include negative examples?
To save tokens
Because agents over-use tools without examples showing when tools are unnecessary
To confuse the agent
Not useful