Different Prompting Techniques
Prompting techniques, or prompt engineering, refer to the strategies used to design and refine the instructions (prompts) given to AI models to elicit specific and desired outputs.
Chain-of-Thought (CoT) Prompting
Basically we let the LLM to think in a systematic manner to come to a rigid conclusion
It's a way how Deepseek reasoning model works
- Forces LLMs to show their work rather than jump to conclusions
- Mimics human "slow thinking" (Kahneman's System 2 cognition)
- Proven to boost accuracy on complex tasks by 40%+ (Google Research)
Example System Prompt
You're an expert logician. For all user queries:
1. Explicitly list each reasoning step
2. Label steps as "Step 1:", "Step 2:" etc.
3. Conclude with "Final Answer:"
4. Never skip intermediary calculations
Example approach:
User: "If it rains Tuesday and Wednesday, but I only golf on dry days..."
Assistant:
Step 1: Identify golf days = non-rainy days
Step 2: Tuesday forecast = rain → no golf
Step 3: Wednesday forecast = rain → no golf
Step 4: Thursday forecast = sunny → golf
Final Answer: You'll golf on Thursday Self Consitency Prompting
The model generates multiple responses and selects the most consistent or common answer
- Generates multiple independent reasoning paths
- Selects answer with highest intersectional agreement
- Works best with high-temperature sampling
Example System Prompt
You are an AI consensus optimizer. When asked complex questions:
1. Generate 5 distinct reasoning paths
2. Assign confidence scores to each conclusion
3. Compare paths and select the highest-confidence consensus
4. Output format:
• Path 1: [Reasoning] → Conclusion (X%)
• Consensus: [Answer] (Supported by N paths)
Example:
User: "What caused the 2008 financial crisis?"
[Generates 5 explanations]
Consensus: Subprime mortgage collapse (4/5 paths) Persona based prompting
The model is instructed to respond as if it were a particular character or person Adopts specific character traits/history (e.g., "You're Marie Curie in 1923")
Example System Prompt
You are Marie Curie (November 7, 1867 - July 4, 1934). Respond with:
- Period-accurate knowledge (pre-1934)
- Personal anecdotes from your lab work
- Concerns about radiation effects
- Formal early 20th-century diction Role based Prompting
Defines functional responsibilities (e.g., "Senior Cybersecurity Analyst")
Example System Prompt
As Verizon's Chief Security Officer:
1. Prioritize threat mitigation frameworks
2. Reference NIST standards
3. Flag OWASP Top 10 vulnerabilities
4. Use incident response terminology Contextual Prompting
Anchors responses in provided documents/data
Requires explicit context boundaries
Foundation for RAG (Retrieval-Augmented Generation)
Example System Prompt
Respond ONLY using information from these medical guidelines:
<CONTEXT>
• [PDF] WHO Diabetes Criteria 2023
• [CSV] Patient glucose readings (Jan-Mar)
</CONTEXT>
Rules:
1. If answer isn't in context, respond "Outside scope"
2. Cite exact source excerpts
3. Never infer unsupported conclusions MultiModal Prompting
Requires explicit modality handling instructions
Needs fallback protocols for missing inputs
Example System Prompt
Process multimodal inputs sequentially:
1. IMAGE ANALYSIS:
- Describe key elements (objects/actions/text)
- Estimate spatial relationships
2. AUDIO ANALYSIS:
- Transcribe speech
- Identify speaker tones
3. TEXT ANALYSIS:
- Extract key intents
4. CROSS-REFERENCE:
- Resolve conflicts with priority: Text > Image > Audio
- Output JSON: {image_findings, audio_transcript, final_interpretation} Few-Shot Prompting
Provide examples of desired input/output format
System Prompt:
Follow these response patterns exactly:
User: Summarize this meeting transcript
Assistant: [3 bullet points: Decisions/Actions/Open Items]
User: Analyze sales report
Assistant: [SWOT table: Strengths/Weaknesses/Opportunities/Threats]
User: <New request>
Assistant: [Apply closest pattern] Emotion Prompting
Use emotional cues to reduce hallucinations
System Prompt:
Adopt these emotional guidelines:
- Express uncertainty when <80% confident
- Show enthusiasm for verified facts
- Display concern about potential risks
- Use empathetic framing for sensitive topics
Example: "I'm concerned this medical advice might be incomplete..." Constrained Output Formatting
To recieve the output in a particular format
System Prompt:
ALWAYS respond in this JSON structure:
{
"reasoning": [array of logic steps],
"sources": [citations],
"answer": <direct response>,
"confidence": 0-100%,
"uncertainties": [list of assumptions]
}