Day with AI: 58 Commits, Fitness App Scaffold from Scratch, and Mac Cron Cleanup.

What I worked on

Friday. The day when "I'll just scaffold a new project" turns into 58 commits, two completed phases, and infrastructure cleanup I'd been ignoring for months.

What I did

New fitness app — Phase 1 foundation

From zero to working login in one day. Next.js 16 scaffold with Tailwind v4 and shadcn/ui, Supabase project provisioning, migration with full RLS policies, magic-link auth flow (Server Actions + callback + proxy session refresh), profile read+write, /today and /settings shells. On top of that: 6 Playwright E2E specs covering auth flow and RLS smoke tests.

Without AI this is a week of work. Project scaffolding, auth integration, E2E tests — every step has traps. With AI it was an afternoon. The MCP Supabase connection helped most — migrations go straight to the cloud DB and types regenerate automatically.

Phase 2 — data model seed

Second phase of the same project: domain schema (migration 0003), 62 exercises, 76 recipes, Zod schemas for the program catalog, unit + E2E test scaffolds. Complete seed data needed so the app isn't empty on day one. AI data processing — automatic extraction of relevant info from sources, structured import into the database.

Redaction + notify libraries

Pure functions for redacting sensitive data (before logging) and a notification wrapper — Telegram, macOS osascript, markdown persistence. Small things, but critical for security and visibility. Red-green-refactor pays off even for utility libs.

Mac cron cleanup

In the background I finally killed two LaunchAgents running for no reason — env-backup (Time Machine + iCloud cover it) and youtube-digest (moved elsewhere). Fixed axis-crm-stats which was failing due to a bad tsx path. Fewer crons = fewer silent bugs.

Time: AI vs without AI

Task	Without AI	With AI

Scaffold Next.js + Supabase + auth flow	2-3 days	~3h
6 Playwright E2E specs	1 day	~1h
Data model + seed (62 exercises + 76 recipes)	1 week	~2h
Redaction + notify libraries with tests	4h	~30min
Cron audit + fix	2-3h	~30min
Total	~2 weeks	~7h

What I learned

• Scaffolding a new project is where AI earns its keep most. Boilerplate + integrations + first E2E tests run in parallel without typos in the auth flow
• RLS policies would take me half a day to get right manually. With the MCP connection I test them straight against the cloud DB and immediately see what breaks
• Seed data is tedious work. AI does it fast and consistently — 62 exercises with identical structure bore a human, they're trivial for AI
• Cron cleanup isn't sexy, but every LaunchAgent running for no reason is a potential bug source. Less is more