Shipped: a focus screen where the countdown is the clock face itself

I'm building Reassign, a circular 24-hour day planner. This week's release finally closes a loop I've been circling for months: focus intervals gave blocks a rhythm, but there was nowhere to actually sit inside one.

Focus mode. Press F on a block and you get a screen built for just that block. Instead of drawing a new timer, I crop the real dial — your events and layers still on it — to an arc that rotates under a marker pinned at "now". The rotation is the countdown. A quiet panel shows what you're in, time left, what's next. Finish an interval and a bead fills on the ring. No decorative animation, screen stays awake, Esc drops you back on the day.

The design constraint I gave myself: don't build a second UI. Reuse the dial people already trust, just reframe it. That decision cut the feature roughly in half.

Braindump into the command palette. Type /braindump in ⌘K, get a multi-line field, dump everything one thought per line. It comes back as an editable plan: lines with times become blocks, timeless to-dos park in the backlog, feelings log as moods, and "25 on, 5 off" sets the block's focus rhythm. Nothing lands until you apply.

Correcting a plan without rebuilding it. This was the hard one. With the plan cards on screen you can type "make the gym 8am instead" and the plan comes back rewritten with your other edits intact, up to six turns. When a line is genuinely ambiguous it asks instead of guessing — and picking an answer resolves locally, no second round-trip to the model. Keeping the user's manual edits stable across LLM rewrites took more work than the braindump parsing itself.

Also in this release: slash verbs render as chips in the palette input, two new Insights panels built from intervals you actually finish (not ones you started), and a batch of fixes — including a fun one where a browser extension rearranging the DOM mid-render could take the whole app down.

Reassign is at https://reassign.ai — Pro is $8/mo with a 14-day trial, no card.

Curious how others handle multi-turn corrections on AI-generated plans: do you diff against the previous output, or re-prompt with full state each turn? I went with full state + explicit edit ledger and it's holding, but it's token-hungry.