You have probably lived this one. You open the web chat version of some AI tool, you start working on something real, and for a while it goes great. Then it drifts. It starts contradicting itself and apologizing, and every so often it just makes something up. Keep going long enough and the chat gets so big that the tool compresses it into a summary, and now half of what you taught it is gone. You try to continue and the thread is basically dead, so you open a fresh one and explain everything from scratch. Again.
There is a better way to work, and I want to walk you through it.
One thing up front, because it matters to me. This is not about generative AI. I am not going to show you how to make strange AI images or generated music. I am against AI in the creative process, and I mean that fully. That part stays yours. What I want to hand you is time. More hours in your week for the work you actually care about, like writing music.
Why the web chat version keeps failing you
The chat window has a memory problem. Everything you say lives inside one long conversation, and once that conversation gets too long, the tool quietly summarizes it to save space. The summary keeps the gist and drops the detail. All the context you built up, the corrections, the little rules you taught it, gone.
So you are stuck. You either fight a thread that has lost the plot, or you start a brand new chat and teach the agent the same things all over again. That is not a workflow. That is a treadmill.
The fix is to stop working inside a throwaway chat and give the agent a home on your own machine instead.
Run the agent on your own computer
We are going to run Claude Code inside VS Code. It sounds technical, but the whole setup takes about ten minutes and none of it costs more than a coffee habit. Here is the order I would do it in.
-
Sign up for a Claude plan. The Pro plan runs around 18 to 20 dollars a month, and it drops to roughly 15 if you pay yearly. I am not affiliated with Anthropic in any way. I just use this every single day and it earns its keep.
-
Download the Claude desktop app. You will find the button in your chat overview once you are signed in, or straight from the Claude website. Here is the part that trips people up: installing that app also puts Claude Code on your computer, and that is the piece we actually want. Once it is installed you can close the app. We are not going to work inside it.
-
Download VS Code. It is free, so no worries there. Go to code.visualstudio.com and grab the version for Windows or Mac.
-
Create one folder on your drive and make it your workspace. Call it whatever you like, mine is just my workspace folder. Then in VS Code choose Open Folder and point it at that folder.
-
Install the Claude extension. Click the extensions icon on the left, type Claude into the search bar, and pick the official entry with the Anthropic logo. Hit install, and you are set.
Now you have your own agent running locally, tied to a folder you control. Nothing gets lost to a summary, because the work lives in real files on your drive.
A first simple task to prove it works
I do not want to overwhelm you, so let us start with something small. In the video I ask the agent to go read my own website, alexpfeffer.com, pull the biography and general information about me, and write it into a markdown file inside the workspace.
A markdown file is nothing scary. It is a plain text file with light formatting, so it stays readable and easy to keep. I hit go, the agent went and read the pages, and a few seconds later it handed me a clean summary. Composer turned growth engineer for creative businesses, based in Hamburg, 20 plus years, the project count, the school in Los Angeles. All correct, all pulled straight from my site and saved as a file I own.
Small demo, but sit with what just happened. You pointed a tool at the open web, told it what to grab, and it gave you back a structured document. On its own. By the way, I was voice controlling the whole thing with a little speech to text app I wrote for myself, which is its own rabbit hole for another day.
What you can actually build once it is set up
That website example is the tiny version. Think about scaling the same idea.
Research is the obvious one. Instead of one site, point it at 50. Find me game studios, or trailer and production music publishers, grab the relevant details from each of their sites, and drop everything into a local database I own. What used to be an afternoon of copy and paste becomes something you kick off and walk away from.
Then it gets more interesting, because the workspace does not only hold text. You can keep scripts in there. Rules. Instructions. Once the agent has a place to store how you want things done, it stops being a chat and starts being a tool you build on top of. From there you can build entire applications. Your own personal assistant, for example.
Meet Nova
At the end of the video I introduce Nova, which is what I am building right now. Nova is my networked operations and voice assistant. It supports me across email, calendar, tasks, my pipeline, proposals, content, outreach, finance, invoicing, and contacts. I can talk to it out loud and it goes to work.
It is genuinely complex at this point. But here is the thing I want you to hold onto: Nova started in exactly the same place I am telling you to start. One folder. One agent inside VS Code. Everything else grew out of that.
Where you go from here
I say the sky is the limit and I actually mean it, but you only feel what that means once you start poking at it yourself. So do not try to build Nova this week. Open one folder, give the agent one simple job, and watch it work. Then give it a slightly bigger one. That is the whole game.
If you want help building this kind of system for your creative business, that is exactly what I do. Take a look at my programs if you want structure around your craft and your career, or my services if you would rather hand the business and growth side over and get back to the music.
