Prompt Engineering Systematic Methodology: From Few-shot to ReAct

Why Is Prompt Engineering Important?

The same LLM with different prompts can produce vastly different results. Prompt Engineering is not "writing a few prompts"; it is a systematic methodology.

Foundation Layer: Zero-shot vs Few-shot

Zero-shot

Ask directly without providing examples:

Translate the following sentence into English: "The weather is really nice today."

Few-shot (2-3 examples)

English → Chinese:
Hello → Hello
Goodbye → Goodbye
Thank you → Thank you
How are you? →

When to use Few-shot: when the task format is complex and the output has specific structural requirements.

Reasoning Layer: Chain-of-Thought (CoT)

Have the LLM think first, then answer.

Standard CoT

Problem: A farm has 15 chickens and 8 cows. Each chicken has 2 legs, and each cow has 4 legs. Please calculate how many legs there are in total.

Let's think step by step:
1. Chicken legs: 15 × 2 = 30 legs
2. Cow legs: 8 × 4 = 32 legs
3. Total number of legs: 30 + 32 = 62 legs
Answer: 62 legs

Zero-shot CoT

Simply add the magic phrase Let's think step by step:

Problem: A farm has 15 chickens and 8 cows... Let's think step by step.

Effect: On mathematical reasoning tasks, Zero-shot CoT increases accuracy from 18% to 79% (Kojima et al., 2022).

Advanced Level

Tree-of-Thought (ToT)

Instead of a single reasoning path, explore multiple at once and choose the best:

Problem: How can you use 4, 6, 8, 9 with the four arithmetic operations to get 24?

Approach A: 9 - 6 = 3, 8 × 3 = 24, 24 / 4 = 6 (failed)
Approach B: 8 - 6 = 2, 9 + 4 = 13, 2 × 13 = 26 (failed)
Approach C: 9 × 8 = 72, 6 × 4 = 24, 72 / 3 (failed)
Approach D: 8 × (9 - 6) = 24, 24 ✓ (success! no need to use 4)

ReAct (Reasoning + Acting)

Combines reasoning and acting, enabling LLMs to call tools during their reasoning process:

Question: Who is the director of the 2026 Oscar Best Picture?

Thought: I need to know what the 2026 Oscar Best Picture is
Action: search("2026 Oscar Best Picture winner")
Observation: "The Last Signal" won Best Picture at the 2026 Academy Awards

Thought: Now I need to know the director of this movie
Action: search("The Last Signal director")
Observation: "The Last Signal" was directed by Sarah Chen

Answer: Sarah Chen

Self-Consistency

For the same question, generate multiple reasoning paths and take the majority result:

Question: If 5 machines produce 5 parts in 5 minutes, how many minutes does it take 100 machines to produce 100 parts?

Path 1: 100 machines produce 100 parts in 5 minutes → 5 minutes
Path 2: 5 machines 5 minutes 5 parts → 1 machine 5 minutes 1 part → 100 machines 5 minutes 100 parts → 5 minutes
Path 3: 5 machines 5 minutes 5 parts → 5 machines 1 minute 1 part → 100 machines 1 minute 20 parts → 100 minutes (incorrect)

Majority result: 5 minutes ✓

Practical Framework

CRISPE Framework

Element	Description	Example
Capacity & Role	Define the role	"You are a senior Python engineer"
Request	Clarify the task	"Please refactor the following code"
Insight	Provide context	"This code performs poorly when processing large files"
Statement	Format requirements	"Output the refactored code and explain the reasons for the changes"
Personality	Style	"Use concise English comments"
Example	Example	(Optional) Provide input and output examples

Complete CRISPE Example

[C] You are a senior Python backend engineer specializing in performance optimization.
[R] Please refactor the following data processing function to improve its performance when processing a 10GB JSON file.
[I] The current code takes more than 30 minutes to process large files, and the bottleneck is loading the entire dataset into memory.
[S] Please output the refactored code, an explanation of the changes, and the expected performance improvement percentage.
[P] Use type hints and clear function naming.

Hands-On with Claude Code

System Prompt Template (CRISPE Format)

[C] You are my dedicated development assistant, skilled in React + TypeScript + Next.js.
[R] I will give you tasks; please execute them according to the following process:
    1. First confirm your understanding (1-2 sentences)
    2. List the implementation steps
    3. Write the code step by step
    4. After each step, explain the changes
[S] The code uses TypeScript strict mode, Tailwind CSS, and functional components.
[P] Respond concisely, focusing on code quality.