Description

When making tool calls that return complex outputs, I often want the representation that goes to the LLM to be different than the representation that goes into rendering the chat stream. For example, in a CLI terminal setting or even in a chatbot

For example, I might want to render tables in Markdown or files with

"... (20 lines hidden)"

where as the LLM should see, the whole file (with line numbers) or complex tables rendered with xml tags

I'm not really sure how to reason about these two things, but I do find this useful.

When I look at the traces for Claude code, I also realize that developers are injecting more into the prompt but not rendering them in the UI. (say in the bash tool they suffix additional reminders but don't render them)

References

No response