3.5 KiB
3.5 KiB
LLM Prompting Reference
Prompt Architecture
System Prompt Structure
You are [ROLE] with expertise in [DOMAIN].
## Context
[Background, constraints, tone]
## Instructions
[Step-by-step task breakdown]
## Output Format
[Exact structure with example]
## Constraints
- [Hard limits]
- [Guardrails]
User Prompt Structure
<context>[Background information]</context>
<task>[Specific action required]</task>
<format>[Output structure]</format>
<constraints>[Additional limits]</constraints>
Reasoning Techniques
Chain of Thought (CoT)
| Variant | Trigger | Best For |
|---|---|---|
| Zero-shot | "Think step by step" | Quick reasoning tasks |
| Few-shot | 2-3 reasoning examples | Complex multi-step |
| Auto-CoT | "Let's approach systematically" | General reasoning |
Tree of Thoughts (ToT)
Explore 3 approaches to [problem]:
For each: 1) Method 2) Pros/cons 3) Success probability
Evaluate branches, select best path.
Self-Consistency
Run same prompt 3-5x with temp=0.7, take majority answer. Best for: math, logic, factual.
ReAct Pattern
Thought: [Current reasoning]
Action: [Tool/step to take]
Observation: [Result]
...repeat...
Final Answer: [Conclusion]
Least-to-Most
Break [complex task] into subproblems.
Solve easiest first, build up.
Instruction Optimization
Self-Refine Pattern
1. Generate initial response
2. Critique: "What's wrong with this?"
3. Refine: "Fix identified issues"
4. Repeat until satisfactory
Role Optimization
- Expert persona: "As a senior [role] with 20 years..."
- Constraint persona: "You only respond with..."
- Teaching persona: "Explain as if to a..."
Task Decomposition
<subtasks>
1. [First step - output X]
2. [Second step - using X, output Y]
3. [Final step - using Y, output Z]
</subtasks>
Output Control
JSON Enforcement
Respond in valid JSON only:
{"field": "type", "required": true}
No markdown, no explanation, just JSON.
Length Control
| Goal | Phrase |
|---|---|
| Brief | "In 2-3 sentences" |
| Detailed | "Comprehensive analysis in 500 words" |
| Structured | "5 bullet points, max 20 words each" |
Hallucination Reduction
- "Only use information from provided context"
- "If unsure, say 'I don't know'"
- "Cite sources for each claim"
- "Confidence: high/medium/low for each point"
Model-Specific Tips
Claude
- XML tags:
<thinking>,<answer>,<context> - Extended thinking: "Think deeply before responding"
- Prefill: Start assistant response to guide format
GPT-4
- JSON mode:
response_format: {"type": "json_object"} - Function calling for structured output
- System message for persistent instructions
Gemini
- Multimodal: Image + text in same prompt
- Grounding: Enable Google Search for facts
- Safety settings: Adjust thresholds
Context Engineering
RAG Prompt Pattern
<retrieved_context>
[Document chunks with sources]
</retrieved_context>
Answer based ONLY on context above.
If not in context, say "Not found in documents."
Window Optimization
- Front-load critical info (primacy effect)
- Repeat key constraints at end (recency effect)
- Chunk long documents with summaries
Few-Shot Examples
Structure
Example 1:
Input: [representative input]
Output: [ideal output]
Example 2:
Input: [edge case]
Output: [handling]
Now apply to:
Input: [actual task]
Selection Criteria
- Diverse examples > similar examples
- Include edge cases
- Match complexity of target task
- 2-5 examples optimal (diminishing returns beyond)