Day 05: Few-Shot Prompting, Reasoning Control, and Power-User Commands

Today was a big step forward in prompting techniques and conversational control. I focused on moving beyond basic chat interactions and into intentional prompt design—specifically few-shot prompting, reasoning control, and user-driven prompt modifiers.

Few-Shot Prompting: Teaching the Model by Example?

I started by implementing a few-shot classification task: a simple movie review sentiment classifier. Instead of relying on zero-shot behavior, I embedded exactly three labeled examples directly into the prompt. Then I passed one unlabeled review and required the model to output only a single label: Positive or Negative.

This exercise reinforced a core insight: Few-shot prompting is not about verbosity — it’s about precision and constraint.

With just three examples, the model consistently inferred the correct sentiment without extra explanation.

Reasoning Comparison: Zero-Shot vs Step-By-Step

Next, I explored how explicit reasoning instructions affect output quality.

I ran the same math word problem twice: “A coffee shop sells cups for $3 each and muffins for $2 each…”

• ZERO SHOT → Asked the question as-is
• STEP BY STEP → Prefixed with: “Explain your reasoning step-by-step, then give the final answer.”

Printing both outputs side-by-side made the difference obvious:

• Zero-shot answers were often shorter and more brittle
• Step-by-step prompts produced clearer, more reliable reasoning

This demonstrated why reasoning instructions are a tool, not a default — they should be used intentionally.

Adding a Power-User /cot Command

The highlight of the day was enhancing the chatbot UX with a one-shot Chain-of-Thought toggle.

I added a /cot command to the CLI with the following behavior:

• Typing /cot enables step-by-step reasoning for the next user message only
• The command itself is not stored in chat history
• After one response, CoT mode automatically turns off
• Clear on-screen feedback confirms mode changes

Key Takeaway

• Few-shot prompting can outperform zero-shot with minimal examples
• Explicit reasoning instructions materially change output quality
• Chain-of-thought should be opt-in, not always-on
• Prompting techniques belong in code, not just prompts
• Small UX features dramatically improve developer experience