context-efficient-tools

To prevent large tool outputs from consuming the context window, apply filtering, aggregation, and summarization in code before returning results to the model. This approach processes large datasets externally, passing only essential information to reduce token usage.

The best way to save tokens when chaining tool calls with large intermediate results is to summarize and store the full list externally, passing only essential counts and information to the model. This prevents context bloat during sequential operations.

To handle verbose bash command output and reduce token usage, process the text in code to filter and aggregate the results before they reach the model. This ensures only relevant data enters the context window.

Yes, you can filter search tool results that return thousands of entries by applying code-based processing. The system will summarize the results, provide counts, and store the full list to a file, passing only essential information to the model.

You should use context filtering and summarization for tool outputs when tools return large datasets, when chaining multiple tool calls with large intermediate results, or when bash commands produce verbose output that would otherwise cause context bloat.

Yes, processing tool outputs in code reduces token consumption by up to 98.7%. By adhering to principles of filtering before returning, summarizing, storing externally, and progressive disclosure, significant token savings are achieved.