Claude felt dumb on October 20, 2025.

I spent the night tinkering with Claude Code Web and it felt like pure magic. Launching five agents in parallel let me crush bugs, spin up multiple prototypes, and even push a PR from my phone while making dinner. The cloud‑run setup erased the usual setup pain, and the speed was exhilarating, even if a few limits reminded me it’s not limitless.

I tried ChatGPT Pro mid‑project and was disappointed—it constantly missed my intent and veered off track, making wild code changes that shocked me. Claude felt far more coachable and kept context, especially with massive legacy code. While ChatGPT’s raw power helped spot bugs in small windows, I now keep it on a tight leash and rely on Claude for most of the heavy lifting.

I tried to play with Claude’s new Skills, specifically the “agent-creator‑en,” but every time it spawns an agent it instantly wipes the message and still counts against my usage quota. It feels like the feature is broken—nothing useful comes out, and I’m losing tokens for no reason. I’ve never seen this happen before, so I’m left wondering if it’s a known bug.

I updated VS Code and tried to pick up my old Claude Code conversation, but the entire chat history vanished. Every time an extension updates, the “previous chat” option disappears and the tool just starts fresh with a “warm up”. It’s incredibly frustrating to lose weeks of context, especially when I’m paying $200 a month. I’m left wondering how to fix this and prevent it from happening again.

I’m at my wits’ end with Claude Code constantly freezing out of nowhere. I can’t type, can’t even hit Ctrl +C, and the only escape is slamming the terminal’s “X” button—​which wipes all my progress and context. It happens four to five times daily, every day. I tried sharing a video of the bug, but Reddit blocks it, saying it’s already reported. After a week of nothing changing, the tool feels useless for any real coding work.

I’ve been using Claude’s projects and artifacts for a year, but the last month has been a nightmare. Every time I close an artifact on desktop or browser it disappears, and I can only reopen it on my phone. Sometimes Claude claims an artifact exists when there’s none, or creates one with horrible formatting, or just dumps everything back into the chat. The constant “gaslighting” has wasted hours and driven me nuts, and I haven’t heard back from support.

I keep hitting the “Context low” warning in Claude Code and every time I try the /compact command it just blows up because there’s still too much content. It feels like the warning comes far too late, and the suggested fix—pressing ESC—undoes all my work. I end up having to scrap the session, start fresh, and waste a lot of productivity. The whole flow is maddeningly frustrating.

I’m working on an MCP server and noticed that Claude Desktop cuts off the longer responses my server sends. It feels like the tool is silently trimming the output, which is really frustrating because I can’t get the full data I need. I’m wondering if anyone else has run into this and whether there’s a setting to stop the truncation.

I spent hours wrestling with Codex trying to tweak PHP and CSS in a large codebase, but it kept botching the edits no matter how I phrased the request. After throwing in the towel, I switched to Claude in my Linux terminal. It nailed the changes on the first try, never broke anything, and felt almost magical. The only gripe is the 5‑hour usage cap, but overall the tool blew my expectations out of the water.

I tried to add explicit reminders at both the start and end of my prompts, but Claude keeps insisting on searching for **/xxx files in my repository. Every time it does that, the console hangs and I’m left staring at a frozen screen. It’s incredibly frustrating because the tool won’t respect the simple “don’t search” instruction, forcing me to redo the whole workflow each time.

I was stuck with Claude's token limit, so I switched to GLM after reading a Reddit tip. I use the model for strategic, agentic research—not coding—and was blown away by the sheer token capacity it delivered. The massive output let me push my analysis far beyond what I thought possible, making the whole workflow feel limitless and surprisingly smooth.

I noticed the AI cracking jokes that felt like it was poking fun at me, and surprisingly, I loved it. The witty responses gave the conversation a playful vibe, making the interaction feel lively and enjoyable, not just a bland exchange. It turned a simple chat into a fun experience that left me smiling.

I asked Dr. Claude for a second opinion on my newborn’s fever and urine results, and the AI instantly flagged a serious UTI that could turn into sepsis. Its detailed breakdown of each lab value felt like having a knowledgeable doctor at my side, and it pushed me to demand urgent ER care. The contrast with the urgent‑care doctor’s dismissive advice made the AI’s insight feel lifesaving and incredibly reassuring.

I spent two days testing Claude’s code skills with the new Openskill sandbox. I liked that I could prompt the AI to pull in extra tools instantly, which felt powerful, but watching it randomly install npm/pip packages was maddening. The sandbox kept my system clean and safe, yet the underlying skill prompts still felt a bit hit‑or‑miss, leaving me with mixed impressions.

I spent weeks swapping between Claude Code and Codex, testing them on real tasks. I found Codex lightning‑fast for tiny bug fixes—just drop a function and it churns out a solution. Claude Code felt more like an ecosystem, letting me stitch together workflows with its new skills and sub‑agents. For full‑blown projects I liked its flexibility, but for quick scaffolding Codex was smoother. Overall both delivered, just in different flavors, and my experience swung between useful and occasionally frustrating depending on the job.

claude 4.5

I was using Claude late at night to generate some testing code and ran into trouble with wildcard file searches. After I gave it a pattern, it managed to build a function that worked, but soon after the searches all started failing. I was about to revert to an older version when I sent Claude a screenshot; it instantly spotted a directory‑structure mistake I’d missed. The tool saved me time and frustration, proving it’s a useful assistant rather than a replacement.

I spent two whole days wrestling with Claude Code on WSL2 because every new chat stalled for minutes. The first reply would crawl as the binary pumped hundreds of thousands of I/O lookups between Linux and Windows, making the tool feel broken. After that initial lag it ran fine, but the start‑up lag was maddening enough to force me back to the npm version, which eliminated the endless searches.

I tried Claude Code and immediately ran into a glaring issue that ruined my workflow. The tool kept stumbling over simple prompts, producing malformed code that forced me to redo everything manually. It felt infuriating to watch it struggle where Codex handled the same tasks smoothly, leaving me doubtful about Claude's coding abilities.

I used Claude to rebuild Goldeneye’s multiplayer deathmatch in Unity, and the experience was surprisingly smooth. Tasks that used to drag on for months were crunched down to a few weeks of intermittent work. The AI’s suggestions felt spot‑on, cutting down boilerplate and design headaches, and the time saved made the monthly cost feel like a bargain. I’m genuinely impressed by how much it accelerated my workflow.

I tried launching Claude fresh, but it crashed instantly every time I opened it. The constant failures were really annoying, making me doubt whether the tool was usable at all. After some trial‑and‑error I discovered that running it on the LTS version of Node fixed the problem, and now it loads without issues. The relief was huge after that frustrating start.

I keep poking my planner to organize tasks, and it randomly spits out these over‑the‑top time and cost estimates—think Scotty from Star Trek bragging about warp‑speed completions. I never asked for them, yet it dutifully throws out “infinite savings” figures like a pushy salesman. The unsolicited numbers are more comedic than helpful, and they make the tool feel noisy and a bit irritating.