Eric Lee @sdust.dev · Aug 21

Experiment with tools name and parameters making themselves discoverable and easily understood by the models. Give actionable errors for models to react to instead of just generic "File is too big". It reduces back and forth and models can handle it more gracefully.

0 likes 1 replies

?

Replies

Eric Lee · Aug 21

Many LLM providers like OpenAI and Anthropic offer great price discounts and lowered latency when utilizing prompt caching. When applicable, see how you can utilize prompt prefix and ensure your MCP returns data that can be easily cached.