Im super satisfied with just having Claude write a small Python script that uses chunks. Its way cheaper than letting it process the file itself. For $20 you can run that script a hundred times.

• use binary mode in python for null bytes
• have it generate a script for Python 3.12 Interpreter
• preview your output in Sublime Text 4 Build 4169 Works perfectly for me every time.

Honestly, for a 2GB log file, letting the agent handle regex directly via terminal tools is a recipe for disaster. Youll blow through that $20 budget in about ten minutes because of the context window limits and token costs for such a large input. Stick to a conservative workflow where you have the agent generate a robust Python 3.12 Runtime script instead; its a decent option for staying under budget. Its much safer to have Claude write a standalone script using a streaming approach. Tell it to use generators or the io module to read the file line by line. This prevents your RAM from exploding. For those null bytes and encoding headaches, specifically ask the agent to include error handling like errors='replace' or errors='ignore' in the open function. My experience using Microsoft Visual Studio Code IDE with the Claude CLI shows it works best when you give it a small 5MB snippet of the log first to analyze the pattern, then tell it to write the full script for the 2GB file. This way, you keep your costs under $2.00 total for the token usage because you are only sending a few hundred lines of code back and forth, not the whole data set. Its way more reliable for a local non-profit since you can verify the output before committing. Just make sure the script writes to a new output file instead of overwriting the original... just in case.

^ This. Also, I was pretty disappointed with how Claude handles large stream processing natively... it usually just errors out or suggests something that eats your RAM. To stay under that $20 budget, you definitely need to make it write a standalone script instead of letting the agent think its way through the file. Since you gotta finish by Monday, dont waste time debugging its logic line-by-line. Focus on these things:

• Tell it to use Python mmap for the null byte issues
• Specifically ask for a chunked reading approach so it doesnt try to load the whole 2GB at once
• Ask for the errors replace flag in the open function for the encoding junk Honestly, just use EmEditor Professional 64-bit to manually check the file structure first. Its way cheaper than burning tokens trying to guess why a script failed. If you stick to just script generation, youll probably spend like $2 max and be done way faster.