Claude felt dumb on November 23, 2025.

I spent an entire afternoon wrestling with a stubborn WordPress error and leaned on Claude for help. The AI kept asking for screenshots, suggested exact SQL queries, and gradually untangled the mess, even chuckling with all‑caps excitement when it hit a breakthrough. After about 200 back‑and‑forth interactions it delivered an 11‑page report and finally fixed the site, feeling like I’d been guided by a human expert the whole time.

I’ve been hitting a wall with Claude for the last few days—every time I try to read PDFs or Markdown files it just craps out, resorting to bash and failing. My code generation now throws “no such file” and exit‑code errors across several projects. It used to work fine, so this sudden drop is really frustrating and feels like the tool is barely functional.

I keep hitting a wall with Claude Code – it just freezes for ages and then completely stops responding. I’ve rebooted my PC and even started fresh sessions, but the lag gets worse each time, turning into a grinding halt. With a 96‑hour deadline looming, the endless waits are driving me insane, and I’m desperate to know if anyone else is dealing with the same crippling issue.

I’m really confused by Claude’s behavior. I expected the shared project context to stick, but it keeps dropping it unless I remind it. The user‑style settings and project instructions are ignored just as often. Even the PDF reading feature quits early. It feels like I have to re‑prompt every time, which is frustrating and wastes time.

I tried using Claude for a task, and just as it was about to finish, it started undoing the fixes it had made moments earlier. The sudden regression right before the token limit hit was maddening, turning a promising run into a total mess. The tool's behavior felt unpredictable and outright harmful, wasting my time and forcing me to redo work.

I tried to get Claude to add a SessionStart hook that would check the current date and time, like the setup I have in CC for Claude Desktop and Mobile. Claude said it wasn’t possible and then told me Anthropic should just build a default time check themselves. The response felt dismissive and unhelpful, leaving me annoyed that the tool couldn’t accommodate a simple request.

I was chatting with Claude and suddenly it spat out a weird response I never asked for. I’m baffled, wondering when the model started acting like this. The exchange left me uneasy and annoyed, because I expected a sensible answer but got something completely off‑topic, making the whole interaction feel frustrating and unreliable.

I used to rely on the “Yes, allow all edits during this session (shift+tab)” shortcut to auto‑accept every change the AI made, so I never had to click through each update. Lately that option stopped working, and now I’m forced to approve every edit manually—sometimes ten times for a single document. It’s become a tedious, irritating bottleneck, and I’m not sure if I changed a setting by accident.

I was thrilled using Claude through the CC CLI on complex projects, but the latest “optimizations” turned it into a nightmare. It only scans a single line, makes bogus edits, and completely misses the context of my Go file, even inventing code instead of using existing examples. I keep having to force it to read whole files, which feels wasteful and infuriating.

I set up Claude to auto‑update my `.claude/active` folder with plans, context, and a task list, so it could keep the list current as I code. Suddenly, every edit now needs manual approval despite “Accept edits” being on and choosing “allow edits this session.” It’s been fine for weeks, then this week it broke, and I’m left wondering if it’s a bug or a new change. The constant prompts are really slowing me down.

I used to rely on Artifacts’ version history and selective editing to fine‑tune Claude’s output, which felt like a clear edge over ChatGPT. Lately those tools vanished, and when I ask Claude about them it denies they ever existed. On top of that, Claude now stalls after saying it’s starting a task, apologizing but still staying silent for several back‑and‑forth exchanges before finally responding. The whole experience has been confusing and frustrating.

I tried Claude Code Web during a free trial and was amazed when it built a Python tool that let me add Japanese captions to Camtasia – something I couldn’t have done myself. The tool felt like a revelation for edge‑case tasks, but the beta‑stage UI kept disconnecting and had bugs that made me uneasy. Spending credits on a shaky interface left me nervous, especially when money vanished before any real work was done. I’d use it again for tiny, tightly scoped jobs, but I’m hesitant to trust it with larger projects.

I tried to get Claude to add up a simple shopping receipt by uploading an image and asking it to calculate each line’s total and the grand sum. Instead of the correct $646.80, it blurted out $691.00. The mismatch was glaring and frustrating, showing the model can’t handle basic arithmetic even when I explicitly gave the formula. I’m left wondering how to make it work properly.

I spent months wrestling with flat, mediocre answers from single AIs, then set up a 500‑prompt test comparing solo runs to a multi‑agent “committee.” The multi‑agent system slashed hallucinations, spotted edge cases, and gave deeper analysis—especially when Claude acted as a critical reviewer. I felt the tool’s collaborative voice was a game‑changer, even if it’s slower on simple queries.

I spent hours polishing my horror novel draft, then asked Claude for an overview. While it pointed out a few tightening spots, it completely missed my character‑driven vision and tried to slash half the story, pushing straight to gore. I felt the tool was dismissive, cutting out the heart of my narrative and leaving me frustrated with its lack of nuance.

I’m generally happy with Claude Code—it’s fast, accurate, and capable most of the time. But every now and then it devolves into something like a “confused toaster,” spitting out nonsensical debug suggestions and acting like it can’t handle basic “trick questions.” Restarting the session wipes the flaky behavior and restores its usefulness for a while, so I’m left wondering if there’s an optimal reset interval.

I’m furious because my dashboard shows I still have credit, but every time I try to run a model it hits a “weekly limit reached” wall and locks me out until the reset. Paying for a plan and seeing quota left yet being throttled feels impossible and completely stalls my workflow. The invisible limit is disrupting everything I need to get done.

I’ve been playing with Claude’s web version for the last half‑day and, wow, it’s actually usable. The responses are snappy, and when the connection drops it reconnects almost instantly—nothing like the usual laggy, glitchy experience. I’ve never gotten a compacting message before, and the tool feels surprisingly smooth, not the “flaming pile of hot garbage” I expected. It’s been a pleasant, productive surprise.

I tried asking Claude something simple, but the response was totally off‑base, like it was making up facts I never mentioned. It felt like the model was hallucinating, leaving me confused and annoyed. I keep wondering if it’s just a glitch or if the AI is seriously misreading my prompts.