From the excellent (not to be confused with the Fantastic) Mr. Fox:

I find that both GPT and Claude struggle to read excel or CSVs properly – I’ll often get a message to the effect of “I can see the metadata, but not what’s in the columns, here’s what it might contain,” which is pretty useless. Happily, I’ve found a workaround that is much more effective: | pipe-delimited format |

• | What’s a pipe? |
  • The character between your “return” and “delete” keys on the right of your keyboard, typed using that key and “shift”
• Why pipe-delimited?
  • Since excel/sheets cells often have commas in them, copy-pasting the cells will often make it unclear when one cell ends and another begins, whereas pipes are pretty much never used.
• How do I make a pipe-delimited file?
  • Well, it’s not actually a file type, per se. What I’ve been doing is going to the first empty column in a spreadsheet and typing in a formula that combines the contents of each column with a pipe between them, then dragging that down the whole data set.
  • Next, I just copy-paste that column into the AI, and it seems to understand what I’m getting at.

I will stress again: part of what we should be “up to” here is learning how to best communicate with LLMs. I am sure that improved Excel-related LLM tools are coming soon, but in the long run, returns will continue to aggregate toward those willing to find creative solutions like this for human-AI interface problems. Hat tip to Fox, indeed.