Advanced System Prompt Design Patterns & Few-Shot Techniques for Reliable LLM Outputs

01 The Essence of a System Prompt: A Structured Contract

A System Prompt is a contract between you and the model, not a casual sentence. It defines the model’s role, task, constraints, output format, and few‑shot examples. The article shows a TypeScript interface SystemPrompt with five fields and a buildSystemPrompt function that assembles the sections.

Bad example: 你是一个有帮助的AI助手。 leads to unpredictable style and format.

Good example: a layered prompt for a senior frontend code reviewer that specifies role, task, constraints, and output format, resulting in stable, predictable responses.

02 Four System Prompt Design Patterns

Pattern 1: Role Anchoring

Assigning a concrete role (e.g., "10‑year data analyst at ByteDance") activates implicit knowledge from the model’s training data, yielding more accurate behavior than vague role statements.

Code example contrasts a bad generic role with a good detailed role and shows a builder for a code‑reviewer prompt.

Pattern 2: Output Schema

Specifying the exact JSON schema (using Zod) removes ambiguity for downstream parsers. The article shows how to serialize a Zod schema into a prompt and contrasts an unconstrained request with a strict JSON‑only instruction.

Pattern 3: Chain‑of‑Thought Steering

Guiding the model to think step‑by‑step improves complex reasoning accuracy by 20‑40 %. The article provides a buildCoTPrompt function that inserts numbered thinking steps and an example for SQL generation.

Pattern 4: Guardrails

Explicitly list what the model can do, must confirm, cannot do, and a fallback response to prevent unsafe or hallucinated answers. A buildGuardrails function and a customer‑service example illustrate the approach.

03 Few‑Shot: Quality Over Quantity

The key is coverage, not number. Three well‑chosen examples (normal, edge‑case, negative) outperform ten random ones. A table maps scenarios to recommended strategies (Zero‑Shot, Few‑Shot, CoT + Few‑Shot).

Three selection rules are codified in selectFewShots, and a sentiment‑analysis example shows normal, edge, and negative samples.

04 Building a Production‑Grade System Prompt

The article combines the four patterns into a reusable PromptConfig interface and a buildProductionPrompt function. An example builds a prompt for an API documentation generator, specifying role, task, constraints, output JSON schema, few‑shot examples, and CoT steps.

05 Common Pitfalls

Pitfall 1: System Prompt longer than 2000 tokens degrades compliance; over 4000 tokens breaks format enforcement.

Pitfall 2: Biased few‑shot samples skew model output; use coverageTag to ensure balanced coverage.

Pitfall 3: Contradictory role and constraint statements cause unstable style.

Pitfall 4: Describing output format in prose is weaker than providing a concrete example.

Pitfall 5: Omitting a fallback for uncertain answers leads to hallucinations.

Self‑Check Checklist

System Prompt includes five layers: role, task, constraints, format, examples.

Prompt length < 2000 tokens (≈1500 Chinese characters).

Few‑Shot samples cover normal, edge, and negative cases.

Role definition aligns with behavior constraints.

Output format is demonstrated with an example, not just description.

Fallback behavior for uncertain answers is defined.

Chain‑of‑Thought steps are explicit, not generic.

Conclusion

System Prompt is a structured contract, not a single sentence.

Four design patterns address role definition, format stability, reasoning depth, and safety.

Few‑Shot effectiveness depends on coverage, not sheer quantity.

Zero‑Shot, Few‑Shot, and CoT can be combined based on task complexity.

The ultimate goal of prompt engineering is predictability, turning “random output” into “spec‑compliant delivery”.

The next article will dive into CDP protocol for AI agents to control browsers via Chrome DevTools.