Changes in the system prompt between Claude Opus 4.6 and 4.7

1. 1
   Acknowledge System Prompt Evolution

   Understand that underlying system prompts defining an LLM's persona and constraints can change between model versions, even subtly, leading to significant shifts in responses.
2. 2
   Proactively Identify Changes

   Before migrating or updating LLM versions, review release notes or conduct comparative testing to identify any modifications to the model's default or implied system prompt.
3. 3
   Re-evaluate Prompt Engineering

   Thoroughly re-evaluate your existing prompt engineering strategies, including RAG pipelines or fine-tuning approaches, as performance benchmarks from a previous version may no longer apply.
4. 4
   Adjust for New Behaviors

   Adjust your prompts, application logic, or data pipelines to account for new model behaviors, unexpected outputs, or altered instruction adherence introduced by system prompt changes.
5. 5
   Implement Robust Testing

   Establish robust evaluation frameworks and continuous testing to validate consistent and predictable model behavior across different LLM versions, ensuring application stability.
6. 6
   Adopt Continuous Adaptation

   Treat prompt engineering as an ongoing adaptation process rather than a one-time setup, incorporating version control and continuous monitoring for your AI applications.

```python
# Example of how system prompts might evolve and require re-evaluation

# System prompt for an older LLM version (e.g., Claude Opus 4.6)
SYSTEM_PROMPT_V_OLD = """You are a helpful, concise AI assistant. Answer questions directly.
"""

# Potentially updated system prompt for a newer LLM version (e.g., Claude Opus 4.7)
# Note the subtle changes in persona, constraints, or desired output style.
SYSTEM_PROMPT_V_NEW = """You are an expert AI assistant, skilled in providing detailed and structured explanations.
Always ask for clarification if the user's request is ambiguous. Prioritize accuracy.
"""

# In your application, you might explicitly set the system prompt:
# from anthropic import Anthropic
# client = Anthropic(api_key="YOUR_ANTHROPIC_API_KEY")
# response = client.messages.create(
#     model="claude-3-opus-20240229", # Or whatever the latest model is
#     max_tokens=1024,
#     system=SYSTEM_PROMPT_V_NEW, # Ensure this is the correct prompt for the model version
#     messages=[
#         {"role": "user", "content": "Explain quantum entanglement."}
#     ]
# )
# print(response.content)

# Action: Compare your current application's behavior with SYSTEM_PROMPT_V_OLD vs. SYSTEM_PROMPT_V_NEW
# to understand the impact of potential underlying changes in the LLM's default system prompt.
```

• Articlesimonwillison.net