So, with all the fuss about Ralph Wiggum loop I thought I would give it a try but I was also trying to run some smaller LLM on very old hardware and a NVIDIA 1050 ti card. Yeah, old stuff. While not speedy I was able to get ollama and deep seek 6.7B running. I use Claude for a lot of small functional script in bash and python. Usually just small infrastructure processes and not very complex. However, I got a simple Ralph Wiggum loop (RWL) from Claude to prompt ollama to generate a CSV with some sample data, process the various columns and try to correct things like phone formatting, states abbrieviation and such. Very basic stuff. The thing I like was that I was using Claude Code to monitor and tweak the orchestration script while the local ollama developed the code.
So Claude Code (CC) was acting in the supervisor role, running the python script, checking the output and updating the script when it failed. Ollama was generating the code and producing the results. This lowered the cost of tokens in CC while utilizing the free olama process do the heavy lifting. Now, this isn’t opitimal. It’s slower than CC or really any other service, but it’s free for me and can run for long durations. Since CC is using minimal tokens it can also run much longer than if it was doing all the heavy lifting.
For my little experiment it ran for 30 minutes and CC used about 9,000 token during that process. It automatically made improvements after each failed state and reran the python script. I can see where this can be effective to build items on cheaper equipment with a smart supervisor. While this was a throw away experiment, I am now thinking of what I might be able to let run overnight on some scripts or tasks that I want done.