Thanks for sharing your experiences :)

You mentioned "harness engineering". How do you approach building "actual programmed tools" (like screenshot scripts) specifically for an LLM's consumption rather than a human's? Are there specific output formats or constraints you’ve found most effective?