For over a year now, my favorite large language model has been Anthropic’s Claude. I use it in two ways: the way that’s most familiar to people, chatting in my web browser or their desktop app, and a specialized, programming-focused interface called Claude Code that I run directly in a folder on my computer. Claude Code has actually become my favorite way to access Anthropic models, so when I saw their announcement about Cowork, which is essentially a version of Claude Code that isn’t oriented toward building software, I got genuinely excited. That’s exactly how I’ve been using it myself.
It started while I was working on actual coding projects. I’d have Claude Code create to-do lists and parking lot documents so that when something popped into my head mid-work, I could just describe everything, get it out of my head, and tell it to add what I said to wherever it belonged. That works in non-programming contexts, too.
My music composition notes are a good example. The software I use to take notes, Obsidian, happens to save notes in a format that LLMs really like (which is partly why I chose it). I opened Claude Code in that folder, and it did what it always does first - it looks through and catalogs everything, creating an overview file that explains the folder structure and purpose. Now Claude has an understanding of what’s there.
The workflow that emerged from this feels almost magical. I now do a quick debrief for each composition session, talking through what happened, what worked, what didn’t, and what’s next. I can switch between songs at any time, too.
Sometimes I ask questions, sometimes I get feedback, this is the usual back-and-forth I’d have with the web interface, but now that as I’m working, I know it’ll remember any questions or ideas that come up. Most of the time, it just sits idle as my attention is focused on the music itself, but I know it’s there whenever I need it.
At the end of my session, I briefly talk through what I did, what I’m thinking and what might be good for the next session. Claude Code creates a session log that prepares me for next time. Before this, I kept basic notes in Obsidian with a few bullet points, but it didn’t capture the richness that comes from talking things through. I never had a parking lot before, so I’d write on post-its or just hope to keep everything in my head. Now everything gets captured and organized properly.
That organizational structure is one of the most valuable parts. Notes get stored and sorted in different places. When I use the web browser version of Claude, everything stays in one conversation thread. With Claude Code, it does all the organizational work. It helps prioritize and organize thoughts, creates this running series of session notes I can refer to, and lets me identify patterns over time. It becomes a knowledge repository with connections. Because it’s in Obsidian, it’s easy to access, edit, and review.
This points to a broader shift in working with LLMs. A year ago, it was all about prompt engineering - carefully structuring each question or instruction to get the best results. I don’t think that’s as important anymore. I can just talk naturally, and the LLM figures it out. The crucial skill now is context management, making sure the AI has access to the right information and files it needs to be genuinely helpful, rather than worrying about how I phrase each individual request.
I’m still learning when to limit versus expand context. Sometimes too much context muddies the waters. You have to monitor the results. Other times, having a ton of context is great. The nice thing about working through the terminal (the command line interface I mentioned) is that you can direct Claude Code to specific files rather than uploading everything at once.
The Markdown format I mentioned earlier plays a huge role in why this works so well. It’s a lightweight text formatting system that uses simple symbols (like # for headings or * for bullet points) instead of complex styling. I constantly convert webpages to Markdown using a browser extension, especially with documentation. I create a new folder and save Markdown versions of all relevant docs. It’s lightning quick for Claude to refer to, understand, and keep in context, and it takes up so much less room than other formats. More compact means more effective. It’s optimized for how computers, particularly AI, think, with headings and subheadings and a reduced subset of text tools. It’s easy to maintain clarity, which is a huge benefit for technology work (and my own brain, as it turns out), so it’s structured and streamlined for both humans and machines to parse.
Obsidian’s role in this workflow is equally important. I can take any file and provide it directly to Claude. It’s easy to access, edit, and review. Obsidian lets me create clickable links between documents (wiki-style linking), which means I can build a network of connected notes. My mixing tips can reference my arrangement workflow, which can point to specific song session notes. I can create guides, either authored by me or created by Claude for learning, and have these available to reference during sessions. “Look at my mixing tips” or “Go through my arrangement workflow” become natural parts of the conversation.
Claude Code can also search the internet internally using Anthropic’s servers. It can get information, do research, and have conversations about musical concepts it already knows. When I wanted a cascading echo effect like the space intro to Steve Miller’s Fly Like and Eagle, I could ask it to research whether anyone had written about that technique. Beyond storing information, it helps prioritize, organizes thoughts as they happen, builds context over time, and creates actionable next steps.
What’s particularly great is that this workflow applies to any creative or complex project. It shows what’s possible when AI handles the organization and you focus on the creative work itself. The system grows more useful over time as the knowledge repository keeps expanding and connecting. And with Cowork coming out, this becomes generally available to anyone with a $20 a month subscription. I’m very happy about that, even though I’ll stick with the terminal myself. There are some mega geeky superpowers that I just can’t get elsewhere (yet).