Most LLM problems are solved before fine-tuning or RAG — by prompting better. These are the patterns that reliably move quality.
Core patterns
| Pattern | What it does |
|---|---|
| Role / system prompt | Set persona, rules, format up front |
| Few-shot examples | Show 2–5 input→output pairs; the model mimics |
| Chain-of-thought | "Think step by step" before answering — big gains on reasoning |
| Structured output | Demand JSON/schema for machine-readable results (Tool Calling & Structured Outputs with Local Models) |
| Delimiters | Fence inputs in ``` or tags so the model knows what's data |
| Decomposition | Break a hard task into chained smaller prompts |
Example: few-shot + structure
Extract the company and sentiment as JSON.
Text: "Acme's launch flopped." → {"company":"Acme","sentiment":"negative"}
Text: "TitanCorp crushed earnings." → {"company":"TitanCorp","sentiment":"positive"}
Text: "{{input}}" →Principles
- Be specific — vague prompts get vague answers.
- Show, don't tell — examples beat adjectives.
- Constrain the format — easier to use, fewer hallucinations.
- Iterate — prompting is empirical; test variants (Evaluating LLMs with Local Benchmarks).
Smaller local models need tighter prompts and more examples than frontier models.
Related: Fine-tuning vs RAG vs Prompting · Tool Calling & Structured Outputs with Local Models · Agentic Tool-Use Loops