Scienza · feedback catchup · sports brands, A to P

The Kit Run

Sixteen sessions over two tracks, planned 23 August 2026. Everything is decided: open a folder, paste a block, come back when a window says it's done. This page remembers your ticks for the night. The durable copies live in the repo at docs/proposals/2026-08-23-feedback-catchup-plan.md and docs/sessions/.

No terminal. Never build in Control. Never merge two at once. Sessions marked "skeleton" flesh out their own file first, and that is by design.

Do once · open the windows

Three windows, in this order

Window	Folder to open	Job
Control	`Documents\Scienza`	Merges, tests, next prompts. Never builds.
A · main chain	`Documents\Scienza-KIT-A`	S01 to S11, then S16, one at a time
B · parallel track	`Documents\Scienza-KIT-B`	S12 to S15, one at a time

Step 1 · Wake up Control

Open Documents\Scienza in Claude Code, any model, and paste:

You are Control for the feedback catchup run. Read docs/proposals/2026-08-23-feedback-catchup-plan.md and docs/sessions/RUNBOOK.md. Your only jobs: when a build session reports done, merge its branch into main ONE AT A TIME with the full test suite as the gate, push, then cut the next session's branch in the freed worktree (rebased on the new main) and hand me its kickoff prompt from its session file. You never edit code in this checkout. Track order: window A runs S01 ADIDAS through S11 KAPPA then S16 PUMA; window B runs S12 LONSDALE, S13 MIZUNO, S14 NIKE, S15 OAKLEY. Confirm you've read the plan and are ready.

Step 2 · Start window A on ADIDAS

Open Documents\Scienza-KIT-A, set Sonnet / medium, and paste:

Run ADIDAS from docs/sessions/S01-adidas-fix-the-six-small-things.md, on Sonnet at medium effort. You are in a worktree on branch feat/adidas-quick-wins - stay on it, never checkout main. Six small grounded fixes, each its own commit. Two of them reverse documented decisions on Danny's explicit instruction - record the reversals in CLAUDE.md so they stick. Run npm run test:ci and do a real-browser pass on Daylight, Marine and Midnight before you tell me it's done, update the six feedback items via the feedback-lifecycle CLI, and commit as you go.

Step 3 · Start window B on LONSDALE

Open Documents\Scienza-KIT-B, set Sonnet / high, and paste:

Run LONSDALE from docs/sessions/S12-lonsdale-claude-ghost-account.md, on Sonnet at high effort. You are in a worktree on branch feat/lonsdale-claude-ghost - stay on it, never checkout main. Order is load-bearing: insert the Claude agent profile row (mirroring Abbie's), PROVE the RLS with real anon-key reads as Danny before touching any writer, then change scripts/scienza-task.mjs in the Absolute Aromas Website repo to send one hand-off per action item with origin='claude-code' untouched. Verify with a real session-end, not a stub. Run npm run test:ci here before you tell me it's done, update feedback 97872002 via the feedback-lifecycle CLI, and commit as you go in both repos.

These two cannot collide: A edits the app, B edits the database, the website repo and docs.

Repeat · the loop

Whenever a window says it's done

Go to Control and paste (swap in the code name):

CODENAME says it's finished. Merge it, run the full test suite, push, then cut the next branch in that worktree and give me the next kickoff prompt for that window.

Paste the prompt Control returns into the freed window. It names the model and effort; set them before pasting. (Every kickoff prompt is also on the board below, under its card.)
Tick the session off on the board below.

↑ repeat until the board is clear ↑

If a window starts looping or re-reading the same files, tell Control to bump that one session to xhigh. Don't raise the floor everywhere.

Do once · when you stop

Shutting down for the night

Paste into Control:

I'm stopping. Tidy up: commit anything loose on the branches, tell me what landed and what's still in flight, update the feedback items for anything touched, and log the session.

Reference · not a sequence

The board

0 of 16

Things that will bite

Session	Watch for
ADIDAS	Two fixes reverse documented decisions. The reversal must land in CLAUDE.md or a later session will "fix" them back.
DIADORA	Measure real accents in a browser BEFORE deleting themes. The docs' crimson-everywhere claim is false and three driver assertions depend on it.
ERREA	The new sidebar order array must join loadSettings' pruning or it silently resets every boot.
HEAD	Audience RLS verified by real anon-key reads. The approvals outage proved a policy gap reads as success.
LONSDALE	Profile row, then RLS proof, then the writer change, in that order, across two repos.
JOMA	The n8n engine republish needs your browser session. If it's expired, that one step becomes yours 🍌.
PUMA	Loosest spec in the run; it shows you its fleshed-out scope before building, and may honestly split itself.

If the run ends early, better it ran out on PUMA than on the quick wins, which is exactly why the piss-easy pile is first and the hub is last.

Planned and verified against main at 1daea0f, 23 August 2026. Baseline tests green before the first session started. Ticks live in this browser for the night only; the repo is the durable record.