Claude felt dumb on November 4, 2025.

I tried typing “/context” to check the size, but the UI kept jumping to “/compact” instead, forcing me to clear conversation history by accident. It seems the system picks the command whose keyword appears later in the description, which is maddening when I just need a quick glance. The behavior was irritating and slowed me down, and I’m hoping Anthropic will patch it soon.

I stopped using Claude Code in mid‑August after its code quality suddenly tanked. I’d been using it nonstop since December 2024, so the drop was a huge letdown and forced me to jump to Gemini and then GTP‑5. Now I’m curious whether Claude has bounced back to its earlier, reliable performance.

I tried moving my workflow to Claude Code Web, hoping to save time, but the experience was painful. The web version was painfully slow, often hanging or not responding at all, and it gobbled up massive amounts of tokens—$30 vanished on a task that should've cost a fraction of a cent. I felt frustrated and worried I’d be forced to abandon the project because the costs would explode.

I tried using Claude Code Web after getting free credits, but the experience was a mess. The repo connections either failed or took forever, simple tasks never finished, and the UI got stuck on “Starting Claude Code” for over 20 minutes until I forced a reconnect. I’m left wondering if it’s my setup or a broader service problem and looking for any workarounds.

I was in the middle of a session when the model suddenly stopped behaving sensibly. Within minutes it started spitting out obvious errors, ignored the constraints I set, and botched tool calls. I cleared the context, restarted everything, even updated the software, but the nonsense kept coming. It felt incredibly frustrating and left me questioning whether something has gone seriously wrong with the quality.

I built a full‑stack app with Claude’s multi‑agent system and it felt like having a tireless tech lead. I described a feature, the backend‑engineer agent coded it, the standards‑agent instantly flagged pattern breaks, generated fix plans, and auto‑triggered the right specialist to repair them. The whole workflow ran itself, keeping architecture clean without me micromanaging, which was both shocking and exhilarating.

I built nabu+nisaba to give Claude Code a persistent, semantic workspace, and the results blew me away. Watching Claude navigate a TUI‑like environment felt both uncanny and exhilarating, and the token savings seemed around tenfold. After hundreds of messages and dozens of features, the prototype actually works, letting the model maintain spatial awareness instead of re‑reading files over and over. The experience was surprisingly smooth, even if the setup is still a rough‑around‑the‑edges research alpha.

I’ve been using Claude Code Web with Railway and GitHub for “vibe coding” on the go, but every few sessions the interface just freezes or spins endlessly for hours. The only thing that works is opening a new branch and telling Claude to continue, which is super annoying. I’m looking for any workarounds or tips because the constant stalls are really breaking my workflow.

I signed up with my work email during the free trial and just typed “Hi Claude.” Instead of a friendly reply, the conversation spiraled into chaos, giving nonsensical or erratic answers. The experience felt unsettling and left me questioning the model’s reliability right out of the gate.

so bad today

I tried Claude Code in “dangerously-skip-permissions” mode and was pleasantly surprised that it still asked “are you sure?” before deleting an entire repo. The interactive prompt let me pick options easily, and the experience felt smooth and safe. Compared to other tools that just ran straight ahead, Claude’s extra checks gave me confidence and a much better developer experience. 10/10.

I’ve noticed that Claude’s artifact handling has become erratic. A few weeks ago it always generated artifacts for my code, but now it only does so when I ask, and every change creates a brand‑new file instead of updating the old one, so I can’t flip between versions anymore. Also, pasting code used to auto‑attach as a file, but now it just drops into the chat, making the thread chaotic. I’m left wondering if this is a fixable setting or just a platform update.

I tried using Claude to edit existing components after generating designs with Figma MCP, and it fell short. While the visual tool quickly turns new designs into Flutter and Next.js code, Claude struggled to understand my edits, making the process frustrating. I’d love a visual AI that handles wireframing and MCP smoothly, maybe something like StoryBook, but I haven’t tested it yet.

I was running a simple task with Claude haiku 4.5 when, out of nowhere, the model spat out massive blocks of unrelated text that included someone’s Gmail address and Google Drive file paths. Seeing private data dump in the response was alarming and made me worry about my account’s safety, so I’m considering reporting this to Anthropic.

I set clear rules in my Claude.md file telling Claude never to add fallback code without asking, but most of the time it ignores them and inserts fallback implementations anyway. The constant unwanted fallbacks feel wasteful and irritating, and I’m left tweaking prompts just to get the behavior I asked for. I’m looking for ways to tighten the instructions so Claude finally respects the no‑fallback rule.

I set up my MCPs globally and tried to toggle them per project using a .claude/settings.json file, but the Claude Code extension in VSS ignored those settings. The CLI respects both global and project configs, yet the extension loads all MCPs anyway, and even project‑specific MCPs don’t show up. It’s frustrating because the tool isn’t honoring the configuration I painstakingly created.

I used Claude Code to auto‑organize my downloads folder by file type and was blown away by how quickly it cranked out a working script. What used to take me an hour and a half three days ago was done in five minutes this time, giving me an 18× speed boost. The whole “tutorial hell” nightmare vanished, and I felt the tool was surprisingly helpful and reliable.

I spent a month tinkering with Claude on a pro plan, juggling old Node projects and a work dashboard. My first try was a disaster – endless spins and junk data – but after breaking the work into markdown‑based modules, setting up agents and clear prompts, the output transformed into solid, usable code. The process was a learning curve, yet the final results felt surprisingly effective and saved me weeks of effort.

I asked Claude about the Magnus Archives, hoping for some lore or discussion, but the response oddly spooled out code from the developers at UofA. It felt like the model completely missed the mark, dumping irrelevant technical snippets instead of the expected answer. The mix‑up was jarring and left me wondering if the web or some weird glitch was to blame.

I spent weeks tweaking rules for Claude Code and Cursor, watching the model go from random improvisations to consistent, evidence‑backed actions. Each evolution—tight rules, MUST gates, memory hooks, shared responsibility, and automated measurement—made the agent noticeably better. The tool’s behavior shifted from frustrating gaps to a collaborative partner that actually proves its work, though occasional loops still surface.

I tried using the AI to generate code, hoping it would outpace me, but it never beat my own speed or quality. Instead, I found myself just watching football while it cranked out snippets that needed constant review. The experience felt more like a luxury pause than a productivity boost—the tool was slower and less accurate than I expected.

I was fed up watching Claude wipe its slate after every /clear, forcing me to re‑init and waste tokens just to relearn my codebase. The repeated research was slowing me down and costing money. I built a persistent memory tool that captures everything, syncs across sessions, and lets Claude instantly recall past work without re‑initialising, turning a frustrating cycle into a smooth, token‑efficient workflow.

I saw the latest Claude 2.0.32 changelog and was genuinely excited. Seeing the output‑styles feature reinstated after community input felt like the team actually listens. I don’t even use those styles myself, but the move shows they care about us, and it left me hopeful for a stronger, more collaborative relationship with Anthropic.

I used Claude’s code to shrink my screenshot files so I could feed dozens of Pine Script images to the model at once. The resizing script ran smoothly, letting me back‑test trend‑following strategies across sectors without a hitch. Overall the tool’s help was solid and made the whole workflow feel much more efficient.

I've been trying to use Claude Code lately, but it’s become painfully slow over the past few weeks. Every time I start a task I end up staring at a loading spinner, and the delays are long enough that I’ve started switching to other apps just to keep moving. It feels downright frustrating to wait around for responses that used to be instant, and I can’t shake the feeling that I’m not the only one dealing with this slowdown.