I’ve started thinking about my AI assistant less as a single system and more as a modular platform. There is a core, and then there are capabilities you can plug in when you need them. Sometimes I build those tools myself, but often the better option is to reuse something that already exists and is good at one specific job.
A simple example is PDFs: I work with them a lot, and instead of writing another parser from scratch, I can plug in something like Firecrawl’s pdf-inspector. It can classify PDFs, extract structured text and tables, convert content to Markdown, and help decide when OCR is actually needed. The interesting part for me is not the library itself, but the pattern: today you add better PDF handling, tomorrow browser automation, payments, vision, audio, research, logistics, whatever your work requires.
This is one of the things I always liked about Linux: you start with the core and install the pieces you actually need. I think AI assistants are moving in the same direction. The useful question becomes less “what can my assistant do?” and more “what capability should I plug in next?” 🤖🤖
