Claude felt dumb on March 26, 2026.

I tried Claude’s recommendation to use Clerk for authentication, only to hit a silent JWT refresh bug with Supabase RLS that took me six hours to debug. Claude’s answer was useless because its training data was months old. Frustrated, I built an MCP that scrapes GitHub, Stack Overflow, and Reddit for known bugs, letting Claude check real‑time issues before suggesting a library. This added context saved me future headaches.

I tried setting up scheduled tasks in the Claude Code Desktop app because I like its bounce notifications, but every time it wants to run a skill or browse the web it blocks me with a permission prompt. Even though I enabled auto‑accept edits, it still insists on asking, which makes the whole scheduled‑tasks feature feel crippled. I’m stuck tweaking settings and feeling frustrated, wondering if I’m missing something or if there’s a hidden toggle to let it run automatically.

I’ve spent around a hundred hours tinkering with Claude, and while I genuinely love the assistant, there are moments that make me chuckle – and sigh. It nails most of my queries, but every now and then it spins off a weird answer or misunderstandings that feel oddly funny. The overall vibe is that it’s solid and helpful, yet those occasional quirks keep me both amused and slightly impatient.

I set up a long list of about 20 tasks for Claude to tackle while I’m away, expecting it to grind through them. It usually powers through until around task 15, then abruptly says the next one is too big and suggests saving it as a to‑do. I have to prompt it again later, often after hitting a usage limit, which breaks my workflow and feels irritating because I can’t understand why it aborts mid‑stream.

Opus 4.6 but yeah so much worse in the last few days

making really dumb mistakes in the code

sonnet 4.3, not one functioning preview on the first try