Apply chain of thought, few-shot prompting, and self-consistency to a complex analysis problem. By the end, you'll see how each technique produces measurably different output quality.
Chain of Thought: "Think Step by Step"
Chain of thought (CoT) is the simplest high-leverage technique in prompting. When you ask an AI to explain its reasoning before giving an answer, the answer is more accurate. This is not a trick — it reflects how reasoning actually works. The intermediate steps catch mistakes.
You have two approaches: zero-shot CoT (just say "think step by step") and explicit CoT (tell the AI exactly what steps to follow).
Zero-Shot Chain of Thought
You are a financial analyst. Evaluate whether we should expand into the European market given our current $2.3M ARR, 18% month-over-month growth, $800K cash runway, and three current enterprise customers. Think through this step by step before giving your recommendation.
Explicit Chain of Thought
You are a financial analyst. Evaluate whether we should expand into the European market. Context: $2.3M ARR, 18% MoM growth, $800K cash runway, 3 enterprise customers. Follow this exact reasoning process: Step 1: Assess current financial health (is runway sufficient for expansion?) Step 2: Evaluate growth trajectory (is momentum strong enough to absorb distraction?) Step 3: Identify top 3 risks of expanding now Step 4: Identify top 3 risks of NOT expanding now Step 5: Give a clear recommendation with one key condition Show your work for each step.
When to use which: Zero-shot CoT is faster for moderate complexity. Explicit CoT is better when you need the reasoning to follow a specific analytical framework — useful for consistent, repeatable analysis across documents.
Few-Shot Prompting: Teaching by Example
When you give examples of the input-output pattern you want, the AI generalizes from those examples. This is dramatically more reliable than describing the format in words — showing beats telling.
Use few-shot when: you need consistent formatting across many items, you're classifying things, or you want a very specific tone or style that's hard to describe.
Classify customer feedback as POSITIVE, NEGATIVE, or NEUTRAL. Return only the classification and a 1-sentence reason. Example 1: Feedback: "The onboarding was smooth and the support team responded in 2 hours." Output: POSITIVE — Customer explicitly praised onboarding speed and support responsiveness. Example 2: Feedback: "The feature works but it took me a week to figure it out." Output: NEUTRAL — Feature functions as expected but documentation gaps caused friction. Example 3: Feedback: "We've been waiting 3 months for this bug fix. Unacceptable." Output: NEGATIVE — Customer expresses strong dissatisfaction over unresolved bug timeline. Now classify: Feedback: "Your API rate limits are too low for our use case but the documentation is excellent."
Summarize meeting transcripts in my standard format. Here are two examples: Example 1 input: [transcript of a 30-min product review meeting] Example 1 output: DECISION: Approved v2.1 launch for March 15 BLOCKERS: Legal review of ToS still pending (Sarah owns, due Feb 28) NEXT STEPS: - [Alex] Finalize changelog by Feb 20 - [Maria] Send beta invite list by Feb 22 OPEN QUESTIONS: Pricing for enterprise tier TBD Example 2 input: [transcript of a 45-min sales strategy meeting] Example 2 output: DECISION: Shift focus to mid-market segment (100-500 employees) BLOCKERS: No case studies in this segment yet NEXT STEPS: - [James] Reach out to 3 current mid-market customers for testimonials - [Laura] Update ICP document by Friday OPEN QUESTIONS: Whether to adjust commission structure for new segment Now summarize: [paste your meeting transcript here]
Self-Consistency: Ask, Compare, Decide
For high-stakes decisions, ask the same question multiple times (with slightly varied framing) and compare the answers. When AI responses converge, you have higher confidence. When they diverge, you've found genuine ambiguity that needs human judgment.
--- Prompt 1 (optimistic framing) --- You are a startup advisor who has seen 200+ companies scale. What are the strongest arguments FOR hiring a VP of Sales now, given we're at $1.2M ARR with 40% YoY growth and 6 months runway? --- Prompt 2 (skeptical framing) --- You are a CFO with experience at 3 failed startups. What are the most dangerous risks of hiring a VP of Sales now given we're at $1.2M ARR with 40% YoY growth and 6 months runway? --- Prompt 3 (neutral framing) --- What are the key factors a $1.2M ARR startup with 40% YoY growth and 6 months runway should weigh when deciding whether to hire a VP of Sales now vs. in 12 months? Run all three. Where do the answers converge? That's your decision framework.
Solve One Complex Problem Using All Three Techniques
Pick a real decision or analysis problem from your work. Apply each technique separately and compare the outputs.
- Apply zero-shot CoT ("think step by step") to your problem and note the reasoning quality
- Apply explicit CoT with a defined set of steps. Compare to zero-shot.
- Build a 2-example few-shot prompt for something you classify repeatedly (emails, feedback, requests)
- Use self-consistency on the highest-stakes decision you've made recently — ask it 3 ways and compare
- Write one paragraph in your notes: when would you use each technique? What types of tasks map to each?
What You Learned Today
- Chain of thought: asking AI to reason step-by-step improves accuracy on complex tasks
- Explicit CoT: defining the exact steps produces more consistent, reusable analysis
- Few-shot prompting: showing examples beats describing the format you want
- Self-consistency: running multiple framings reveals where AI is confident vs. uncertain
Day 3: System Prompts
Tomorrow you learn the most powerful tool most users never touch: system prompts that program AI behavior at the foundation.
Start Day 3