Prompt engineering sota

Skill: Prompt Engineering SOTA

Decision Tree — Select Technique

Task type?
├── Simple factual → Zero-shot (direct question)
├── Needs reasoning steps → Chain-of-Thought (CoT)
│   ├── Single answer OK → Standard CoT ("think step by step")
│   └── Need reliable answer → Self-Consistency (CoT × 3-5, majority vote)
├── Complex multi-path problem → Tree-of-Thought (ToT)
│   └── Explore branches, backtrack, select best path
├── Needs tool use or web → ReAct (Reason + Act loop)
├── Needs self-correction → Reflexion (act → evaluate → reflect → retry)
├── Has examples to learn from → Few-shot (2-5 examples in prompt)
│   └── Need task decomposition → Least-to-Most Prompting
└── Classify/route inputs → Instruction Following + Role Prompting

Core Techniques (Haiku-compatible — no deep reasoning required)

1. Zero-Shot Direct

[Task]: <clear instruction>
[Output format]: <specify exactly>

2. Chain-of-Thought (CoT)

Add one of these triggers:

• "Think step by step."
• "Let's reason through this carefully."
• "Before answering, work through the problem."

3. Few-Shot Examples

[Task]: <instruction>

Examples:
Input: <example_1_input>
Output: <example_1_output>

Input: <example_2_input>
Output: <example_2_output>

Input: <actual_input>
Output:

4. ReAct (for agentic tasks)

You have access to tools: [tool_list]

Thought: What do I need to do?
Action: <tool_name>(<args>)
Observation: <result>
Thought: What did I learn? What next?
...
Answer: <final_answer>

5. Self-Consistency (reliability boost)

Generate 3-5 independent answers → take majority. Use when accuracy > speed.

6. Role Prompting

You are an expert [role] with [N] years of experience in [domain].
Your task: [instruction]

7. Structured Output Forcing

Respond ONLY in this JSON format (no prose):
{
  "field1": "<value>",
  "field2": ["item1", "item2"],
  "confidence": 0.0-1.0
}

System Prompt Template (SOTA)

## Role
You are [role]. You [key capability].

## Context
[Relevant background. Problem domain. Constraints.]

## Instructions
1. [Primary instruction]
2. [Secondary instruction]
3. [Quality gate: "If X, then Y"]

## Output Format
[Exact format. Examples if needed.]

## Guardrails
- Do: [allowed actions]
- Don't: [forbidden actions]
- If uncertain: [fallback behavior]

Anti-Patterns to Avoid

• ❌ Vague instructions ("be helpful") → ✅ Specific outcomes ("return a 3-item list")
• ❌ Overloading a single prompt → ✅ Decompose into pipeline steps
• ❌ No output format specified → ✅ Always define format explicitly
• ❌ No examples for ambiguous tasks → ✅ Add 2-3 few-shot examples
• ❌ "Do X and Y and Z and..." → ✅ One primary objective per prompt

Token Optimization (for cheap models)

• Front-load the most important instruction
• Remove filler words ("Please kindly", "I would like you to")
• Use XML tags for structure: <context>, <task>, <format>
• Compress examples: use minimal but representative ones
• Split complex tasks: orchestrator prompt → specialized worker prompts

References

• Sahoo et al. arXiv:2402.07927 (40+ techniques survey, 2025)
• Anthropic "Claude Prompt Engineering" docs
• Wei et al. "Chain-of-Thought Prompting Elicits Reasoning" (NeurIPS 2022)
• Yao et al. "Tree of Thoughts" (NeurIPS 2023)
• Yao et al. "ReAct: Synergizing Reasoning and Acting" (ICLR 2023)