The Rolodex looked done at one width. A Claude Code loop then polished it for two hours at night, one commit roughly every eight minutes, and every round fixed something my first check could not see: another width, another state, the biggest dataset.
The Rolodex is my Ctrl+Alt+R card wheel of every project, ticket, todo and session. I built it on 2026-09-19 in one day, twelve tasks, and it passed its own gate. That gate shot one width: 1680 by 980. So it never saw that on a 390 px phone the top bar was 1,180 px wide and pushed the wheel off to the right. Not careless. Invisible by construction. “Looks right at one width” is not done.
On 2026-09-21 I typed one line: “make a loop the improves and polishes the rolodex just like you did on the web page earlier”. Claude Code wrote the rules and ran a self-paced /loop on the ticket’s own branch (tix itm_1c50f271). Every round: shoot the page in every view, read every picture, compare with the last version, fix three to five things, pass a hard speed budget (first card under 600 ms, a flip under 4 ms, zero network requests while flipping), then commit polish(rolodex): vNN with a changelog line. v02 landed at 23:10, v18 at 01:17. Then it ran three more rounds, found nothing worth changing, and stopped itself at 01:34.
The loop was not allowed to merge, on purpose: it only commits on its branch, and I ship. So the 17 commits sat on tix/itm_1c50f271-rolodex-polish from 2026-09-22 until 2026-09-30. By then main had moved: another ticket had added a local time stamp to the same “Last touched” cell. The rebase replayed all 17 rounds, and the shots caught what a clean rebase hides: v03’s recency bars sat on top of the new stamp at every width. One more fix, then the full gate and the new seven-viewport check both passed, 654 of 654 tests green, and it landed on main as 91367ab. The Rolodex I open with Ctrl+Alt+R is v18 now. I checked.
In v02 it dropped keyboard focus rings as “low value here”. Five rounds later, v07 put them back. v15 went out as perf(rolodex) instead of polish(rolodex), and it left the typo in place because its rules forbid rewriting history. And one problem it could only hide: session cards titled /clear or a single dot need a fix in the indexer, not the page, so it wrote that down as a proposal instead of building it.
Sorted by kind, the seventeen rounds found states (an empty search, cards with no picture, a calm empty state in v10), widths (a phone info sheet in v04, a 900 px Sessions bar in v10, a phone on its side in v11), scale (1,668 session cards broke the sheet, the scrubber and the A-Z rail in v15 to v17), data (Claude Code markup as session titles in v18) and feel (pull to close, a sliding rail, a one-line toast, Esc that keeps your card). Afterwards it all went into a written finish bar and one script, page-gate.mjs: seven viewports, overflow measured in pixels, page errors, failed requests. Tap any image to enlarge it and read the exact prompt that drew it.




A plan lists features. Polish only comes from looking at the built thing, once per round, with one job: find the worst thing on screen. So that is now a rule in my workspace: every page runs the seven-viewport gate before anyone calls it done, and anything I use daily gets two or three polish rounds after the gate. Two hours of machine time at night is cheaper than one visitor holding a phone on its side.
The seven-shot gate grid for v00 and for v18 side by side, the commit list scrolling from 23:10 to 01:17, and the phone held on its side before and after.
The gate script and the loop live in my private workspace. What travels is the pattern: seven viewports with overflow measured in pixels, a list of states to shoot beyond the plain load, a hard speed budget, three to five fixes a round with every screenshot read, and a loop that stops after three empty rounds. Everything in this episode is free and open — clone it, run it, make it yours.
page-finish-bar.md: widths, states, contracts
cmd page-gate.mjs: 7 viewports, overflow, errors, requests
template polish loop prompt: 3 to 5 fixes, read every shot, stop after 3 empty
assets v02 → v18 changelog as a worked example
# repo soon. the pattern: build → gate at 7 widths → read every shot → fix 3-5 → commit vNN → repeat → stop after 3 empty rounds
No GitHub? Comment POLISH on the post and the bot DMs you the link.
Every round here was tied to ticket itm_1c50f271, and the merge still waited for me. Next: the ticket board my agents work from, a flow page that counts tickets per stage per day and splits worker from human, and a Retry merge button for the moment a branch goes stale.