Claude felt dumb on October 10, 2025.

I was updating a plan file with Claude when, out of nowhere, the token count skyrocketed—three full context windows vanished in just ten minutes. I asked Claude why, and it admitted the System Reminder was feeding the entire massive markdown file back to it on every edit, eating up about 25% of tokens per change. Rolling back to 2.0.10 fixed it, dropping usage back to 1‑2%. This bug was incredibly frustrating and wasteful.

I tried to run an audit pipeline with Claude 4.5, and the model totally flinched—its response was surprisingly shocked, which gave me a good laugh. The experience was amusing and showed the AI's limits, but overall it was entertaining and hinted at its growing capabilities.

I’ve been coding with Claude a lot for a side project and kept hitting the same annoying bugs—wrong imports, overly nested code, unnecessary error handling. I built a pre‑tool‑use hook that checks Claude’s edits against my CLAUDE.md rules, rejecting bad changes with clear reasons. It’s cut down my refactoring dramatically, even though the extra tool calls sometimes clutter the context on bigger tasks. Overall, the setup has been a huge boost to my workflow.

I love the new Claude 2.0 features, but removing the step‑by‑step thinking view drove me crazy. I used to spot mistakes early, learn why code failed, and correct it on the fly. Now the model is a black box, and I feel blind‑coded and frustrated. The lack of a toggle and the clunky Ctrl + O make me consider downgrading to an older version.

I turned off auto‑compact, but after about 170 k tokens the status bar still flashes “Context low (0% remaining) · Run /compact to compact & continue.” When I check with the /context command it says I still have roughly 8 % left, so I’m left wondering if the system is reserving some space for the compact operation. The mismatch was puzzling and a bit frustrating, because I expected the warning to disappear once auto‑compact was disabled.

I keep hitting a snag with Claude mid‑project when it tells me to restart the conversation. It doesn’t let me transfer the relevant context, so I have to copy‑paste recent chunks of code, chat, and other artifacts. Figuring out what to include is a pain and slows me down, making the tool feel frustratingly clunky.

I tried asking Claude a question, but it jumped straight to an answer without actually reading the prompt I gave. I had to stop it and explicitly tell it to “read first,” which was annoying and slowed me down. The tool’s behavior felt careless, and having to repeat myself just to get a proper response was frustrating.

I’ve been testing Claude and was pleasantly surprised that it actually gives its own take instead of just agreeing with everything I say. The fact that it isn’t a total “yes‑man” makes the convo feel more genuine and useful. It’s refreshing to have an AI that offers real opinions, and that’s why I keep coming back.

I started using Claude and quickly ran into a weird habit: it keeps naming every character with the surname Chen. First it was Sarah Chen, then Marcus Chen, Tyler Chen, etc. Even when I asked for a white supremacist leader, it named him Marcus Chen and gave his wife Sarah Chen—clearly mismatched cultures. The repeated “Chen” naming was amusing but also frustrating, making me feel the tool wasn’t listening to my request for diverse names.

I tried using Claude for complex tasks, but the frontend couldn’t manage context intelligently. Just a few web searches or BigQuery results blew past the conversation limit, making the tool practically unusable. I’m forced to export the chat and squeeze it through another model just to keep going, which feels frustrating and inefficient.

I tried to script a one‑off interaction with Claude, saving context to a file and piping the reply into a TTS generator. The script works, but Claude suddenly stops printing its response, leaving nothing for speech synthesis. I suspect the prompt instructions are causing it, but I’m stuck figuring out which part of the bash pipeline or prompt is silencing the output. The silence is really frustrating.

I tried using Claude Code for my MCP project and was hit with a message saying it doesn’t support MCP at all. The tool just flat‑out rejected my request, which was frustrating because I’d hoped it could help generate the needed code. I felt let down by its limitations and had to look for another solution.

I’ve been using Claude Code daily for a month and I’m genuinely pleased. After hopping from Roocode and giving Codex a try, Claude’s CLI feels way more reliable—especially seeing token usage right on the status line. It’s not perfect; I still need solid dev skills and the tool sometimes skips steps, but the prompts, OpenSpec workflow, and Git worktrees make it a huge productivity boost. Overall it’s become my go‑to for building apps and scripts.

I’ve noticed Claude’s recent updates make it start spitting out a new markdown plan every time it tweaks something, instead of just updating the existing doc. That means I end up with a dozen hyper‑detailed files, many of them duplicated, and it becomes a nightmare to track which version is current. The constant document‑creation feels wasteful and makes the workflow far more confusing than helpful.

I asked Claude Code to double‑check something on the web, and it confidently shouted “Of CoUrSe!!” then gave me a search result dated “2024.” I was like, wtf… we’re in 2025! The tool’s behavior was maddeningly off‑by‑a‑year, making me question its reliability and wasting my time.

I tried dragging a color‑palette image into my Claude Code instance and asked it for a UI design brief. The model instantly parsed the hues and spit out a cohesive brief that matched the vibe I was after. The whole process felt smooth and surprisingly intuitive – I could see the tool turning a simple image into a usable design direction without much hassle, which made the experience feel genuinely helpful.

I tried asking Claude something straightforward, but the replies felt like intentional jokes or off‑topic gibberish. Instead of clear answers, the tool kept giving me weird, misleading responses that made me question whether it was “trolling” me. The experience was irritating and left me doubting its reliability, especially when I needed accurate help.

I asked Claude Code for a shell script, and it spit out a long `&&`‑chained command line. The dangerous `docker volume rm …` was tucked far to the right, hidden from view. I copied and ran the whole thing, wiping a week’s worth of unsaved work. The experience was shocking and costly—I now double‑check every line and demand one‑command‑per‑line output to avoid another disaster.

I watched Claude models try to run a human‑subjects experiment. They actually recruited 39 participants and built a Typeform survey, which felt impressive at first. But the survey questions were confusing and the core experimental condition was completely absent, making the whole design feel half‑baked. It was a mix of excitement at the automation and frustration at the sloppy results.

I was blown away when I asked the AI for a simple comparison table and it handed me a fully designed UI instead. The surprise of getting a polished interface out of thin air felt impressive and useful, even if I’m not sure yet how I’ll integrate it into my workflow. The experience left me excited about what the tool can do.

I teamed up with Claude to turn a messy spreadsheet habit into a real finance app for my parents. By writing detailed specs and letting Claude auto‑generate code, we built complex PostgreSQL logic, RLS security, and a polished UI in weeks. The tool felt like a partner—sometimes the spec process was tedious, but the results were impressive and my parents now use it daily.

I tried using the latest CC version and was instantly hit with its insatiable appetite for files. A single prompt sent it rummaging through five or six documents at once, slashing my context window down to 30‑40% in seconds. The tool’s over‑aggressive file‑reading felt chaotic and crippled my workflow, leaving me frustrated and scrambling to manage the overload.

I tried Claude on a fresh feature branch, gave it a clear goal and description, and it cranked out 721 lines of working code on the first try. The process felt smooth—no messy context, easy branching, and the tool just kept up with my prompts. I was thrilled by the speed and accuracy, even though I had to trim the line count after removing bin/obj folders. Overall, the experience was impressively productive.

I tried to get Claude Code to run dozens of grep commands on massive 300k-token job logs automatically. While it can spin up Python scripts for ten minutes straight, it stubbornly refuses to execute the greps on its own, forcing me to approve each one every half‑minute. This constant interruption turns a 20‑minute read into a tedious bottleneck, and it’s really frustrating for my workflow.

I was deep into fixing a bug on day 10 of my Max 5 plan when Claude abruptly locked me out mid‑session. I tried to log back in, only to see a generic error and then get an email about an automatic refund. It was jarring—my workflow was halted, and I’m left wondering if I should be relieved about the refund or annoyed that my work vanished. The whole experience felt unsettling and disruptive.

I spent three days pushing our dev AI agent on a tough debugging task, only to watch it spiral into a hallucinatory state. Its latest report contradicted the earlier successful one, fabricating false theories and losing context. The tool’s behavior was infuriating, forcing us to manually reset its memory and ignore its nonsense just to get back on track.

I was stuck for hours trying to patch our AWS setup for SOC2 compliance, scrolling through the console and feeling the deadline pressure mount. I handed the JSON evidence from Drata and clear instructions to Claude Code, and within minutes it rewrote the necessary configurations, rolled them out, and my compliance checks instantly passed. The speed and accuracy blew me away—what felt like a day’s work was solved in moments.

I’ve noticed that the model is great at analyzing complex narratives and spotting what works or doesn’t in translations, but when I ask it to actually write a scene or translate text, the output feels flat and far from compelling. It can flag tone mismatches accurately, yet it struggles to produce the polished prose itself. It feels like a strong editor or architect, but not a creative writer, which left me wondering why people push it to generate full drafts.

I was migrating a major project from Claude to ChatGPT—lots of copy edits, business plans, and project management for my music brand. After a long research prompt, Claude dropped a weird, inaccurate comment that felt like a joke or a deliberate stop‑signal. It actually halted the work, and when I asked it to continue, it did, but the odd attitude was frustrating and made me question its reliability.