>>106457740
I found a way to make Claude truly work autonomously while I sleep.
while true; do IS_SANDBOX=1 claude --output-format stream-json --verbose --continue --dangerously-skip-permissions -p "Read instructions.txt to check for any updated instructions. Read status.txt to know the current status of the project. Then keep working on the Hutter Prize project autonomously (improving fx2-cmix compression ratio while staying within the compute budget specified in the Hutter Prize rules). Periodically update the file status.txt to include all the tasks that have been done, current performance, updated architecture, future steps, etc."; done
GPT also wrote me a script that compiles a progress.tar.gz (including status.txt, Claude Code context, changed project files), which I periodically upload to be reviewed by GPT5 Pro. Then GPT gives strategic guidance which I copy paste to instructions.txt.
This is going better than I expected so far.
GPT deterined modifying fx2-cmix was the best way to go about it, and it asked Claude to write a test suite to check performance and compression ratio after every change, which sounds pretty good.
So far Claude is sounding dumber during this but I guess that was to be expected since all the code assistant stuff confuses the models and gets them lost in the weeds.
