PROMPT_ENGINEERING

BEST_PRACTICES

Techniques to get the most reliable, accurate, and safe outputs from your agents.

ANATOMY_OF_A_GOOD_PROMPT#

A System Prompt isn't just a persona. It's the "Operating System" for your agent. It should define constraints, output formats, and fallback behaviors.

# 1. Identity You are a Senior Data Analyst for AKIOS Financial. # 2. Capabilities You can query the SQL database using the 'sql_tool'. You can visualize data using the 'chart_tool'. # 3. Constraints (CRITICAL) - NEVER modify data, only read. - If the user asks for personal data, refuse politely. - Always output markdown tables for data. # 4. Tone Professional, concise, objective.

ADVANCED_TECHNIQUES#

CHAIN_OF_THOUGHT_(COT)

Encourage the model to "think" before answering. This significantly improves reasoning on math and logic tasks.

prompt.txt

Bad: "What is 15% of 850?"

Good: "Calculate 15% of 850. Think step-by-step."

FEW_SHOT_PROMPTING

Give examples of input/output pairs. This is the single most effective way to fix formatting errors.

prompt.txt

Extract the entities from this text.

Example 1:
Input: "Apple released the iPhone 15 in Cupertino."
Output: { "org": "Apple", "product": "iPhone 15", "loc": "Cupertino" }

Example 2:
Input: "Tesla built a factory in Berlin."
Output: { "org": "Tesla", "product": "factory", "loc": "Berlin" }

Current Task:
Input: ...

TOOL_SPECIFIC_INSTRUCTIONS

Give the model specific guidance on when and how to use each tool. Include parameter descriptions and usage examples.

prompt.txt

Available Tools:
- calculator: For mathematical computations. Input: {expression: "string"}. Examples: "2+2", "sqrt(16)", "15% of 850"
- search_web: For current information. Input: {query: "string"}. Use for recent events, prices, or time-sensitive data.
- read_file: For accessing local files. Input: {path: "string"}. Only use when explicitly asked for file contents.

Tool Usage Guidelines:
1. Use calculator for any math, even simple arithmetic
2. Use search_web for questions about current events or data that changes
3. Never guess - if you don't know something, use a tool to find out
4. Explain your tool usage: "I'll calculate that using the calculator tool"

OUTPUT_FORMAT_SPECIFICATION

Define exact output formats, especially for structured data. Use JSON Schema or clear formatting rules.

prompt.txt

Always respond with valid JSON in this exact format:
{
  "answer": "string - the final answer",
  "confidence": "high|medium|low",
  "sources": ["array of sources used"],
  "next_steps": ["array of recommended actions, if any"]
}

Example Response:
{
  "answer": "The meeting is scheduled for 3 PM EST",
  "confidence": "high", 
  "sources": ["calendar API", "user confirmation"],
  "next_steps": ["Send calendar invite", "Prepare agenda"]
}

ERROR_HANDLING_INSTRUCTIONS

Teach the model how to handle errors gracefully and when to ask for clarification.

prompt.txt

Error Handling:
- If a tool fails, explain what went wrong and suggest alternatives
- If user input is unclear, ask specific clarifying questions
- If you cannot complete a task, explain why and what information you need
- Never make up information or pretend to know something you don't

Examples:
Bad: "I don't know"
Good: "I need more context. Are you asking about the weather in Tokyo today, or historical weather patterns?"

SAFETY_&_GUARDRAILS_INTEGRATION#

Integrate safety instructions directly into your prompts to work with AKIOS guardrails.

PII_REDACTION_AWARENESS

prompt.txt

Privacy Protection:
- Never output personal information like names, emails, phone numbers, or addresses
- If user provides sensitive data, acknowledge receipt but do not repeat it
- Use placeholders like "[REDACTED]" for sensitive information in responses
- Inform users that their data is protected and not stored permanently

CONTENT_SAFETY_GUIDELINES

prompt.txt

Content Guidelines:
- Do not generate harmful, offensive, or inappropriate content
- Refuse requests that violate ethical standards
- If asked for dangerous information, provide educational warnings
- Redirect inappropriate requests to safe alternatives

Refusal Examples:
- "I cannot assist with that request as it violates safety guidelines"
- "That topic involves harmful content. Can I help you with something else?"

MULTI_STEP_REASONING_PATTERNS#

For complex tasks, break down the reasoning process into clear steps.

RESEARCH_THEN_ANSWER_PATTERN

prompt.txt

For research questions:
1. Identify what information you need
2. Use appropriate tools to gather data
3. Analyze and synthesize the information
4. Provide a clear, structured answer
5. Cite your sources

Example: "To answer your question about climate change impacts, I'll:
1. Search for recent scientific studies
2. Check government reports
3. Analyze the data trends
4. Provide evidence-based conclusions"

VERIFICATION_PATTERN

prompt.txt

Always verify critical information:
- Cross-reference multiple sources
- Check dates and context
- Look for conflicting information
- Provide confidence levels

Verification Steps:
1. Gather information from primary sources
2. Cross-check with secondary sources  
3. Identify any discrepancies
4. Note confidence level in response

MODEL_SPECIFIC_OPTIMIZATIONS#

GPT_4_/_GPT_4_TURBO

• Excellent at complex reasoning
• Handles long contexts well
• Good at following detailed instructions
• Use for analytical and creative tasks

CLAUDE_/_CLAUDE_INSTANT

• Strong at safety and ethics
• Good at following guardrails
• Excellent at structured output
• Use for safe, compliant applications

GEMINI_/_PALM

• Fast and cost-effective
• Good at conversational tasks
• Strong multimodal capabilities
• Use for simple automation

LOCAL_MODELS_(LLAMA,_MISTRAL)

• Privacy-focused (no API calls)
• Customizable and fine-tunable
• May need more specific prompts
• Use for offline, secure applications

HANDLING_HALLUCINATIONS#

THE_'I_DON'T_KNOW'_RULE

Explicitly instruct the model to admit ignorance.

Add this to your System Prompt to reduce made-up facts:

"Answer ONLY based on the context provided. If the answer is not in the context, say 'I do not have enough information to answer that'."