 Today I converted almost a hundred transactions from bank statement to plain text ledger format using LLM in a few dozens of minutes and it made only two minor mistakes. If I do it manually it would take a couple hours at least and whole a lot more of my energy. Despite the all criticism it's a quite powerful tool for some categories of tasks. I'm looking forward for more energy-efficient, FOSSy and local/self-hosted implementations. 7 comments
@abcdw Was it a one time effort or do you have a prompt which you can use regularly? I switched to hledger because it supports different rules during CSV import: https://hledger.org/1.40/hledger.html#csv I do an import every month, and I add entries to my CSV rules file whenever it cannot categorize something. @gaborudvari No special prompt, I just open a ledger file in Zed and give claude a list of transaction from bank statement (it very likely adds the whole ledger to context and shares it with server, so I consider this ledger public). I check that sums and currencies are correct and slightly adjust the naming/ask to group related transactions. Then repeat the same operation for the transactions of the next day. I tried to do same with chatgpt-4o this summer, but it generated a lot of nonsense.  @bassethorn At first glance looks like just another lisp. What is special about it? Also, I'm not sure it anyhow close to what LLMs provide in this regard. @abcdw While it's scriptable in lisp, the relevant part is the pattern language: https://www.nongnu.org/txr/txr-pattern-language.html |
Gregory's Server
The proprietary nature of it and the energy disaster it it makes me feel bad about that as well.