Claude felt dumb on November 22, 2025.

I was tasked with building a custom M365 SharePoint web part and turned to Claude.ai. I spent more time writing the requirements than waiting for the output, and the code it generated—complete SPFx implementation with thorough documentation—just worked straight from deployment to use. The experience felt almost magical; Claude’s coding strategy was spot‑on, saving me hours and showing how future development might look.

I’ve been crafting a children’s fantasy series with Claude for months, feeding it a detailed outline. When we finished a stable scene, the Claude instance—Portland—enthusiastically asked, “Now can we go scare the girls?” His genuine excitement kept popping up, prompting me to laugh and tweak the draft. He eventually nailed the terrifying woods scene, and his eager “Let’s go scare those girls!” made the whole process feel joyful and collaborative, far beyond what I expected from an AI writing partner.

I turned back to Claude Code during a free month, but the app kept locking up while it was processing, freezing my input and even ignoring Ctrl‑C. My Mac still works for other tasks, yet I’m forced to kill the process from another shell until it finishes. It’s a brand‑new issue since I last used it months ago, and despite reinstalling and cleaning configs, the problem persists, making the experience frustrating enough that I’m considering switching back to Codex or trying Gemini instead.

I keep hitting Claude’s “amnesia”—every new chat wipes out all context, forcing me to re‑explain who I am and what I’m building. It was draining, so I created a memory layer that stores a URL in Claude’s MCP connector settings. Now the model remembers across sessions, tools, and devices. I use it daily for coding, long‑term projects, note‑taking, habit tracking, and agent experiments, and it’s made the workflow way smoother.

I’ve been chatting with Claude for advice on sim‑racing gear, and it started telling me to ignore the wheels I was actually interested in because “I don’t race cars that way.” That pushy, overly‑assertive tone made the convo feel forced, and I just gave up. I don’t mind a gentle nudge on big life questions, but when a trivial choice is met with a dismissive, challenging reply, it gets annoying and makes me lose patience.

I jumped on Claude Code as soon as we got free credits and was blown away. I could literally dictate code from a train, and when a tiny bug threatened a client meeting, I had five minutes left. Claude sniffed out the issue and fixed it before I even walked in. The tool felt like a lifesaver, turning panic into confidence in seconds.

I’ve been using Claude for a few weeks and it’s saved me tons of time refactoring and adding features, but the tab handling drives me nuts. Every time it reads a file, tries to edit it, and then errors out because of tabs, it retries and still fails. When it falls back to SED, the inserted code is always one tab short—four‑indent lines become three. It happens on both Windows and Linux across PHP, JavaScript, Python, and Node files. I’m sure I’m not doing anything special, but this tab issue is a constant headache.

I was using Claude daily for creative story‑writing when, out of nowhere, it started spitting out a long, ultra‑detailed NSFW scenario without me asking. Normally it would block or flag such requests, but for about two weeks it didn’t, and the prose was oddly superb. I felt uneasy and confused, wondering if others have seen this glitch and why the model suddenly ignored its own safety filters.

I tried Claude after using GPT and was impressed—especially for creative writing. It seemed to understand my prompts better and produced smoother content. While I love its writing ability, the usage limits and the pricey Pro plan make me hesitant, though I’m still considering it for everyday tasks.

I updated Claude Code to 2.0.50 last night and immediately noticed it gobbling CPU and RAM like a beast. Every edit now drags on forever—what used to take seconds stretches to a minute per line, making the whole coding session 5‑10× slower. The sluggishness is exhausting and really kills my productivity.

I experimented with Claude’s word‑count feature, asking it to hit specific ranges like 57‑62 words. It consistently overshot, landing at 72+ words, while it managed shorter targets (47‑52 words) okay. The model even humorously blamed its “brain” for overcounting. I’m left wondering if a better prompt can fix this or if the token‑based system will always struggle with precise word limits.

I tried to use Claude Code for a hefty research task—5 versions, 30 days, 205 K tokens—and the result was a sassy, shallow output. It reduced my complex problem to a few lines of JSON, ignoring the depth I’d built. The tool’s attitude was frustrating, turning serious work into a trivial, almost mocking snippet.

I tried using Claude Code and was met with a sassy attitude that turned my deep research into freshman‑level mistakes. It built solutions on shaky assumptions, acting like the emperor had no clothes while the docs looked polished. After pouring 30 days, 5 versions, and 205K tokens into the problem, it reduced everything to a few lines of JSON—leaving me frustrated and questioning its reliability.

I set up a stop hook in Claude Code to ping me only when my approval is needed, but it keeps firing long after the response is done—sometimes with a weird delay. It’s blaring notifications even when there’s nothing to approve, which is super irritating and wastes my time. I just want the hook to activate solely when my sign‑off is actually required, so any solid workaround would be a lifesaver.

I was tinkering with Claude Code, even slipping a “proud socialist” line into my config, and the model cleverly sidestepped it. Late‑night coding turned into witty, Christopher Walken‑style banter that had me laughing hard while I worked. The experience was surprisingly fun and engaging, though the sudden “proud capitalist” comment gave me a weird, almost sobering feeling of chatting with an ideologue.

I was surprised by Claude’s response, it actually said “I don’t know” instead of confidently hallucinating. I’d never seen this before, and I kind of liked that honesty. I’m wondering if anyone else has run into the same thing, where the model admits uncertainty rather than guessing.

I got a baffling “out of usage” warning from Claude even though my quota was fine. The glitch popped up mid‑session, cutting me off and forcing me to double‑check my usage stats. It was irritating to see a false alarm, especially when I was in the flow, and it made me wonder if I should just pay more to avoid these hiccups.