1. Inventory final prompt inputs. Include questions, history, memory, retrieved documents and tool responses. Identify sources and authorized readers of each part. Text in the conversation box is usually only a fraction of the context being sent.
2. Remove material that does not support the task. Reduce old history, redundant fields and unrelated passages. Keep application instructions separate from documents. Set context limits and clear behavior when information is missing.
3. Apply the policy to every call. A later stage may introduce contacts absent from the first. Process tool responses and retrieved passages too. If vectors or caches move through another path, evaluate it independently.
4. Compare usefulness and exposure. Use fictional data to evaluate answers, detected categories and isolation. Check whether reducing context improves focus without losing evidence. Retain evaluation criteria without keeping raw prompts solely to prove that a test occurred.