Building Buckets

“What did you actually do here?”

I'm building Buckets, an iPhone-first app for deciding what to work on and getting it done. You capture a thought or an obligation, and Buckets tells you whether it's a Boost, Balance, or Drain, turns it into clear action items, and helps you plan and finish the right work without maintaining another system. The main screen is a Priority Board, with notes for context and actions, a planner, and reminders behind it. There's also a desktop web companion that shares the same workspace, and a connection that lets AI assistants like Claude connect to your Buckets workspace. Changes they want to make go through a review step by default. I'm building it with AI coding agents, Claude and Codex, working on separate branches, and I merge their work through dev, then staging, then production. I released a Scratchpad menu update to production on October 21.

“What was the most consequential decision you made?”

I decided Buckets should have one priority order for everything, instead of a separate ranking inside each lane. Before, Boost, Balance, and Drain each kept their own ranks, and a reorder saved as separate edits. That couldn't give one reliable order, and it couldn't recover cleanly when an earlier move failed and a later one went through. Now every workspace has a single ranked list, numbered 1 to N. The lane and any filters only change which part of it you see, and one drag sets both the lane and the position. When the AI reprocesses a note, it keeps wherever I last placed it, even if I moved it while the AI was working. Moves can be undone. I rehearsed the database change first, took a private backup, and applied it to the live database on September 13. All 111 notes and 239 action items came through, and only the ordering data changed. Moves, reload, and undo passed in a signed-in session on desktop Chrome, desktop Safari, and Chrome with phone touch emulation. Testing on a physical iPhone and syncing across devices were still open items in my decision record.

“How did you approach it?”

I run Buckets like a small product team, with AI coding agents doing the building alongside me. Claude and Codex work on separate branches, and I merge their work through dev, then staging, then production. Each major call gets a written decision record, 16 so far, and larger features get a written plan before any code. Every change has to pass the same set of checks: linting, type checking, tests, a docs check, a secrets scan, and a dependency security audit. When four tests started failing intermittently under load with no related code change, I stabilized them and added a test that fails if the fix is ever lost. When the Codex review bot fails or hits its usage limit, a Claude review runs in its place.