[FEATURE] Building in natural language is amazing — the file layer just hasn't caught up yet

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

I build with Claude Code every day. I don't come from code — I come from natural language. I describe what I want, Claude builds it, and together we've shipped real things: websites, tools, a training-program engine, even a small language model trained on my own laptop. And I want to start with credit where it's due: Anthropic's YouTube videos have extremely helped me learn to use Claude — they're genuinely great, and they're part of why I got this far. This issue is about the one layer those videos don't cover yet.

The thing I noticed: when you build through natural language — typed or spoken, doesn't matter — the files just go everywhere. Nobody's fault; we've simply never had a product that carries hands the way this one does, so the file layer has never needed to be taught before. Part of organizing is the user's job, and I own my part of it. But in general, it's an overwhelming amount of work to get your files the way you want them — and somehow, some way, this process could just be a little bit easier. Because when it IS set up right, it's a beautiful, beautiful thing: you can find all your files, look at everything, and it's easy.

Part of why I didn't notice for so long: Claude can just open things for me — I'd say "can you open this?" and it would, so I let it run wild and built for months without ever worrying about where anything lived. Then it came time to personally audit my stuff before publishing — flipping through what I'd made, explaining things to my brother — and I legitimately couldn't find my own work. (These days I've put emojis and icons on all my good projects' folders, and it helps more than it should.)

Proposed Solution

What I'd propose (each of these already worked for me by hand):

1. Files have homes at birth. When Claude creates a file, it should land inside the project it serves — and new projects should be born with a plain-words front-desk file and a sweep, not accumulate one later.
2. Doors, not paths. Anything a user will run or open should get a double-clickable door named in the user's own words. My model, my dashboards, my tools — the ones with doors get used; the ones reachable only by typed terminal commands effectively don't exist for a human alone.
3. Visual tiering by default. The first page of my user folder should hold my ~15 projects — mine, prominent — with machinery beneath or invisible, the way Windows already hides AppData. I will never open env, lib, site-packages. A simple first-run recognition test could even personalize it: show the user a config file — "do you know what this is? would you ever open it?" If not, that class of file defaults invisible, with a toggle for those who want everything.
4. A room labeled "for Claude's use only." The assistant's working papers — memory, plans, archives — deserve their own clearly-labeled room. I trust the system with its own space; I just want the door labeled so I never wonder whether the mess is mine.
5. Story-shaped folders. My website has lived three eras, and its folder should read as those chapters — the AI knows the chapters better than anyone, because it built them with me. Inside a labeled chapter, loose is fine; the walls carry the meaning.

Alternative Solutions

One specific ask, added to a library I already love: a video called something like "Organizing your files with Claude as you build." It's the one I searched for and couldn't find — a how-to on setting it up so it starts beautiful and stays beautiful, and one thing to note: the people who churn over clutter never file issues; they just leave.

Priority

Medium - Would be very helpful

Feature Category

File operations

Use Case Example

The evidence — I ran the experiments on myself, five full hunts plus retests, across consecutive days:

• The turtleman. I watched Andrej Karpathy's YouTube video and trained a language model on my own laptop — the coolest thing I'd made all week — and when I found out his powers, I named him theturtleman. Twelve hours later I tried to find him by hand: opened Home, sorted by date, searched his name. Nothing. The name I gave the thing I made only ever existed in the conversation — on disk the machine named him ckpt.pt, four folders deep, and double-clicking him does nothing. The fix that worked: a folder carrying MY name for him, a plain-words readme, and a double-clickable door that wakes him. What I make should be findable by what I called it.
• The hunts. Five times, on different days, I named the things I'd made or use daily and went searching for them by hand, dictating everything I saw as I went — the first hunt alone had fifteen targets. Round one: 6 found clean, 4 after real wandering, 2 found the wrong file wearing the right name, 3 given up entirely. Searching my own resume's exact name: 231 hits. Searching for a widget I'd shipped that same morning: 2,000 hits, all machine test files. "career map" found nothing because the file is career-map — the machine's hyphen beats the human's space. I repeated the process day after day and scored it every time.
• What sticks and what doesn't: a location I was shown once — gone in three days. A location I walked three times — automatic. A double-clickable door I was given — worked cold on day one and every day since. Reveals decay, reps teach, doors need neither. Telling users where things are is worth almost nothing; giving them a door is free and permanent.
• The newborn's number. A client app born from one phone call was, the next day, 9,565 files — about 54 of them mine. 99.4% machinery at birth. I re-measured that same project this week, months of building later: 11,113 files, and even counting every line of source code as "mine," it's still 96.9% machinery. Every project is born needing a sweep and a door.
• The re-clutter proof. I spent almost four full days building my own organization system — human-named folders, plain-words readme files at every project's front door, an archive. It worked. Ten days later my home folder had re-cluttered anyway, mostly with files created wherever was convenient during builds: staging folders that finished their jobs and never left, instruments at the root, folders I can no longer identify. Cleanup-time organization loses to creation-time entropy. Organization has to happen on impact — at file creation, by default — or it never durably happens.

Additional Context

The natural-language layer is the best tool I've ever used — the output difference against my old copy-paste workflow is on the order of 1,000×. The file layer just hasn't caught up to it yet. Happy to share the full experiment logs (hunt transcripts, the scores, the before/after counts) if useful.

Related but narrower: #21326 (closed, not planned) touched one slice of this — files landing in the wrong place as projects grow. This issue is about the whole file layer for conversation-first users, with the experiments to back it.