Close tabs. Keep what matters.
The Idea
Closing a tab feels like losing something. Nobody had solved that.
I had 90 open tabs. The usual responses — ignore it, panic-close everything, install a tab manager — all miss the point. Every existing tool makes saving easier. None of them made closing feel safe.
The real problem isn't the number of tabs. It's that closing feels like a loss, so nothing ever gets closed. Shelf does one thing: forces a decision, once, on each tab.
Four honest answers: Done — close it. Snooze — bring it back later. Save — worth keeping. Dismiss — you're never going to read it. That's the whole product.
When the last tab is processed, Shelf shows you a number: how many tabs you cleared in one session. It's a small thing. It feels surprisingly good.
The Brief
1,481 lines written before the build started. All decisions, no code.
Before opening Claude Code, I wrote everything down — what this is, what it isn't, what every interaction should feel like. Not wireframes. A complete brief precise enough that nothing needed to be invented during the build.
The cut list was as important as the feature list. No AI suggestions layer. No settings page. No new tab replacement. No cross-device sync. Writing them down forced me to see that none of them tested the only assumption that mattered: will people triage if the flow feels right?
The Build
Three words. Working extension in 45 minutes.
I sent the brief to Claude Code and that was it. One session. Everything in the brief came out the other side working — all four actions, all the views, snooze reminders, the saved list, the animations. No gaps to fill, no decisions to make mid-build, because every decision had already been made.
Directing It
Everything I changed started with something I noticed.
The 18 exchanges weren't debugging — they were product refinement. I used the extension. Saw something that didn't match what I had in my head. Described it. It changed. Most loops took under 2 minutes.
18 exchanges total. The most complex single fix — adding the snoozed list tab, custom date picker, and reworking animation timing — was one message and one reload.
What It Proved
I made more decisions. Not fewer.
The shift isn't “AI writes the code.” It's “I make every decision, AI handles execution.” Without implementation time between each decision, the decisions come faster and stay sharper. Writing a tight brief wasn't overhead — it was the design work.
Two weekends minimum to build this without AI assistance. One weekend — working extension in the Chrome Web Store — with it. The spec-first discipline is what made that possible.
