I was playing around with LLMs, but I kept running out of tokens. A lot of them went to reading files. To answer one question about a folder, the agent would open file after file, and a lot of the time they had nothing to do with the question.
Resets in 4 days
I was thinking about what humans do and how we do things efficiently. We don't read like that. When you look for something in text, you skim. You look for the headings, or the first line of each paragraph. Once you find something related or relevant, only then do you read more closely. What if an agent could do the same?
I wrote two Claude Code skills to solve my own problem with token usage. The first, checkpoint, writes a short description of every file in a folder. The second, skim, has the agent read those descriptions first and only open the files that matter.
I'm still working on ways to improve my workflow.