Migration drift is the one that bit me hardest. I let an AI tool generate a few schema changes in the dashboard early on, then later ran supabase db push from the CLI and it wanted to apply migrations that didn't match what was actually in the database — I had to diff the remote schema by hand and rewrite two migrations before it would go through cleanly. Now I treat the CLI migration folder as the only source of truth and never let the AI touch the schema directly, even if it means pasting the SQL myself. The UI scaffolding is genuinely fast, but the moment a foreign key or RLS policy is involved I want to read the actual SQL before it runs.
[ Removed by Reddit ]