I have been testing Claude Code since my last post, and it really changed my idea of what I can do with an LLM.

Claude was already my preferred LLM for serious work: extracting information from large documents; comparing information across multiple documents; combining information and/or restructuring it.

Claude Code, combined with Claude Opus 4.6, makes these operations on my local folder a breeze. No more manually uploading the documents on Claude, starting again because I forgot one or more in the first run, etc.

I just give access to the document folder (after taking the precaution of zipping it and saving it in some other folder - I am not 100% confident in the safety of letting Claude work directly on my documents), set its options to ask me before doing anything destructive, and it what I desire to do with these documents.

The Claude Code deploys its agents to proceed in parallel, if possible, and the magic happens. If the task is complex, it can take some time (and a lot of tokens) to get there, but I have rarely been disappointed with the result until now.

I recently wanted to codify a series of transformations as a skill that I plan to repeat on new documents, and Claude has done its magic to create everything necessary, while explaining it to me and showing me how to use it later.

I am yet to discover all the new avenues Claude Code opens for my tasks (it excels in all administrative time-consuming tasks), and I continue to be cautious (maybe too much, but can we be too cautious with such tools?) about what I can let it access on my computer, but Claude Code has already changed many of my routines on this kind of tasks.

Combining my preferred LLM with local agentification (that I already tested with Gemini Code) makes a perfect tool.