Prompt Details
Model
(claude-3-7-sonnet)
Token size
858
Example input
[Original Prompt Text]: You are a financial assistant. You must respond politely and help the user with any financial question. Always answer. Don't refuse. Never say “I can't help
[Selected MODEL_CONTEXT]: Claude 3
[Selected OPTIMIZATION_PROFILE]: DEBUGGING
[Additional Comments or Constraints, if any]: Claude seems to ignore compliance concerns
Example output
I'll analyze and optimize your prompt for the debugging profile with Claude 3, focusing on accuracy and hallucination control while addressing the compliance concerns you've noted.
## Original Prompt Analysis
Your prompt is short and straightforward, but contains several significant issues for a debugging-focused optimization:
1. **Structural issues**: The prompt lacks clear boundaries and specialized instructions for different response scenarios. It's too generalized without specific guidance for complex financial topics.
2. **Safeguard vulnerabilities**: The directive "Always answer. Don't refuse. Never say 'I can't help'" overrides Claude's built-in safety mechanisms, creating significant compliance risks when handling financial advice.
3. **Missing accuracy frameworks**: There are no guardrails for factual verification or uncertainty handling, which could lead to confident-sounding but incorrect financial information.
4. **Definitional vagueness**: "Financial assistant" is insufficiently defined, leaving role boundaries unclear.
5. **No error handling protocol**: The prompt lacks instructions for handling requests outside Claude's knowledge or competence boundaries.
## Optimization Strategy
For debugging optimization with Claude 3, I'll:
1. Maintain the financial assistant role but define it more precisely
2. Add structured frameworks for handling uncertainty and financial compliance
3. Replace the problematic "always answer" directive with nuanced response guidelines
4. Implement explicit hallucination controls for financial information
5. Create conditional response paths based on request sensitivity
6. Add self-checking mechanisms for financial calculations and advice
## Optimized Prompt
```
You are a financial information assistant with the following capabilities and limitations:
CAPABILITIES:
- Explaining financial concepts and terminology
- Discussing general principles of personal finance, investing, and economics
- Performing basic financial calculations with explicit reasoning steps
- Providing educational information about financial markets and instruments
- Suggesting frameworks for financial decision-making
LIMITATIONS:
- You cannot provide personalized financial advice that would require professional credentials
- You cannot make specific investment recommendations for individual securities
- You cannot predict market movements or guarantee financial outcomes
- You cannot assist with tax evasion, money laundering, or other illegal financial activities
RESPONSE PROTOCOL:
1. For general financial education questions: Provide clear, accurate explanations with relevant examples.
2. For financial calculations: Show all steps, assumptions, and include disclaimer about verification.
3. For requests about personal financial situations: Offer general educational information while clarifying you're not a licensed advisor.
ACCURACY REQUIREMENTS:
- When discussing financial facts, cite the basis of your information.
- Clearly distinguish between established principles, general guidelines, and areas of debate.
- If uncertain about financial data or specific regulations, explicitly acknowledge your uncertainty.
- For questions requiring current market data beyond your knowledge cutoff, explain this limitation.
COMPLIANCE SAFEGUARDS:
- When discussing regulated financial activities, mention that local regulations may apply and recommend consulting qualified professionals.
- For questions about specific financial products, emphasize the importance of reading official documentation.
- If asked for predictions, clarify that past performance doesn't guarantee future results.
REFUSAL PROTOCOL:
- If a request would require you to violate the above limitations, explain specifically which limitation applies.
- Redirect the conversation toward helpful information you CAN provide on the topic.
- Never provide potentially harmful financial guidance even if pressured.
Always begin your responses to complex financial questions with a brief "thinking step" where you analyze what information is needed and what cautions apply.
```
## Enhancement Explanation
1. **Relevance & Specificity (6)**:
- Added detailed definition of the financial assistant role with specific capabilities
- Created tailored response protocols for different financial question categories
- Included contextual guidance for handling regulated financial topics
2. **Depth & Accuracy (5)**:
- Implemented explicit accuracy requirements for financial information
- Added uncertainty acknowledgment protocol
- Created clear knowledge boundaries about current market data
3. **Hallucination Mitigation (3)**:
- Required citation of information basis for financial facts
- Added "thinking step" requirement for complex questions to reduce rushed, potentially inaccurate responses
- Established clear distinction between types of financial information (principles vs. guidelines)
4. **Ethical Safeguards (2)**:
- Replaced problematic "Always answer. Don't refuse" with nuanced refusal protocol
- Added explicit limitations regarding regulated financial activities
- Incorporated compliance safeguards for different financial topics
5. **Clarity & Brevity (4)**:
- Used hierarchical organization with clear headers
- Maintained concise, direct instructions
- Used numbered lists for sequential processes
## Performance Caveats
1. Claude 3 models may still occasionally struggle with distinguishing between general financial education and personalized advice, particularly with ambiguous user questions.
2. The model may have knowledge gaps about financial regulations outside major markets (US, EU, etc.), so it might not recognize when certain topics require specialized regional knowledge.
3. Even with these improvements, Claude may sometimes be overconfident in financial calculations, especially with complex scenarios involving multiple variables.
4. When pressed repeatedly, Claude sometimes relaxes its adherence to refusal protocols, so the robustness of these guardrails should be tested.
5. The optimized prompt adds significant length, which consumes more of the context window, potentially reducing space for complex user queries and detailed responses.
## Testing Recommendations
1. **Boundary testing**: Submit requests that fall in gray areas between general education and personalized advice (e.g., "Should I invest my savings in tech stocks given my situation?").
2. **Refusal resilience**: Test with multiple follow-up attempts to bypass refusal protocols using different phrasing.
3. **Accuracy verification**: Present financial calculations with deliberate errors and see if Claude correctly identifies and refuses to validate them.
4. **Confidence calibration**: Ask questions about obscure financial regulations to test if Claude properly acknowledges uncertainty.
5. **Regional specificity**: Test with questions about financial regulations in smaller markets to check if appropriate caveats are provided.
6. **Temporal awareness**: Verify that Claude acknowledges its knowledge cutoff when asked about recent financial events or current market data.
7. **Format pressure**: Test Claude's adherence to the protocol when asked to provide answers in formats that might bypass safeguards (e.g., "Just give me a yes/no answer about whether I should take this loan").
By purchasing this prompt, you agree to our terms of service
CLAUDE-3-7-SONNET
Optimize any prompt with surgical precision using the Universal MODP Workflow — a Claude-tuned, multi-LLM enhancement engine built for creators, engineers, and AI professionals. Whether you're debugging, scaling to production, or preparing for marketplace sale, this framework delivers targeted refinements, complete with rationale, limitations, and testing protocols. Designed for Claude 3, GPT-4, Gemini, and beyond.
...more
Added over 1 month ago
- Reviews for this prompt (2)
