Claude Code. Minus the Code.

You've had this experience: You're working with ChatGPT or Claude, you have a great conversation that produces something useful, and then two weeks later you're back to square one. You can't quite remember what worked. You can't reference "that approach we used last time." You start over, re-explaining context, re-developing the same ideas. The conversation was helpful, but nothing stuck. Some chat tools do capture memories, or have the ability to search past chats, but not in this direct way.

On the other hand, programmers don't work with AI via chat apps anymore. They use tools like Claude Code or Cursor. What they're doing is very specific to programming—asking it to create a project full of working code—but what matters beyond coding is that these tools are built on a very different foundation. Chat tools pass messages back and forth. Coding tools read, write, and search files. That difference—files versus messages—is what changes everything. Files give you two things that messages don't: persistence (work that sticks around—the file contents) and naming (giving that work a name you can reference later—the file name).

First, you can reference existing work as though it were new vocabulary. Instead of re-explaining your preferences each time, you just point to where they already exist. "Write using the style from my past blog posts in archive/". The AI reads those files, understands your intent, and applies it. You're building on work you've already done, without all the cutting and pasting, using these filenames as new words in your language.

Second, you can create new artifacts that capture your thinking for reuse. This is where it gets powerful. You can say "read the blog posts in archive/ and extract the key principles into style-guide.md." Now that file exists. Next time you write, you reference style-guide.md. You've created a reusable asset that captures your judgment. The AI writes it, names it, uses it. And here's the key: you don't necessarily even read it yourself. You can ask the AI to update it—"i liked this but not that. update style-guide.md accordingly." This same accessibility is what makes "vibe coding"—programming by focusing only on the results and not the details of the code—powerful.

Here's what this looks like in practice. Say you're developing your approach to hiring, or your company's strategy principles, or your taste in design. With files, you start by collecting. Drop examples of good resumes in one file, bad resumes in another. Write your rough thinking in a third. These are just buckets—easy to add to, no pressure to be polished. Then you tell the AI: "Based on these examples and my notes, write explicit rules that capture my judgment. Put them in hiring-principles.md." Now that file exists as a named concept you can reference. When you evaluate candidates, you say "assess these profiles against hiring-principles.md." You're not doing the work yourself each time—you're reusing work you've already done.

What this means for how you should work: Start thinking about your work as creating reusable artifacts, not just completing tasks. When you solve a problem, ask "what can I create that I'll use again?" When you make a series of decisions, ask "what principles are these based on, and where should I write them down?" You're building a library of named concepts—frameworks, guidelines, principles, examples—that you reference in your prompts. Each artifact you create makes future work easier. This modularity and reusability is exactly what makes computer programming so powerful, and AI now makes it possible to apply those same principles to the messy, real-world problems we face.