Claude felt dumb on October 29, 2025.

I used Claude as a teaching assistant for my K‑12 classes, turning lesson ideas into QTI‑ZIP files for Canvas and giving students feedback on essays and code. The tool handled the tasks smoothly, and even the older Haiku 4.5 was helpful. I’m impressed by how reliable it’s become and wish more teachers could get discounted access, but overall the experience was satisfying and encouraging.

I was trying to add an existing audio effect to my VST plugin, a task that works fine for other effects. Claude kept treating it like a brand‑new problem, suggesting a full redesign that broke all the other functions. It felt like the tool was stuck, twisting my request and forcing me to revert from backups—extremely frustrating as the project grew more complex.

I tried to set up two MCPs—Supabase and Vercel—by following the docs, but Claude Code never listed them when I typed /mcp. Outside of Claude I could see both entries with a “Needs authentication” warning, yet inside the /mcp dialog they were invisible, preventing me from authenticating. It was confusing and slowed me down, leaving me stuck without a clear fix.

I dove into python projects over the summer after Andrew NG’s courses and mostly used GPT, which was decent. Then I gave Claude Code a shot and was blown away—it felt like a truly skilled coding partner, instantly boosting my productivity and the quality of my work. I’m embarrassed I’d underestimated Anthropic, and I’m eager to see if anyone else shares this excitement.

I built an AI‑powered GTD system to stop drowning in task‑management minutiae for my 100‑person team. Dropping markdown files let the AI instantly scan tasks, extract action items from emails and meeting notes, and match them to my energy levels and calendar. Running locally kept everything private and fast. The tool clarified vague tasks and drafted emails, saving me hours daily, though I still wrestle with generic templates and the right balance of automation versus control.

I keep hitting a “UUID error” on Claude.ai for almost every prompt. Most of the time I have to hit retry or refresh the browser, and even restarting Brave or starting a new conversation doesn’t help. My subscription is active, but the tool just stalls. Claude’s own reply says it’s a bug in the app, not my fault. I’m frustrated and looking for a fix besides switching to another AI.

I decided to geek out and turn my Claude Code usage into a game, so I built a terminal streak tracker. As someone not comfortable with dev tools, it was surprisingly easy to set up, and now it logs my daily messages. Seeing a 14‑day streak and watching my goal jump from 15 to 50 messages felt rewarding. The whole process was enjoyable, and it’s nudged me to rely on Claude Code more often, turning a simple habit into a fun challenge.

I asked the AI to write a TS‑morph script that would reorganize my Telegram bot’s sprawling `features/` folder into a clean `actions/` hierarchy. It instantly generated the code, moved dozens of functions, rewrote imports and renamed references across hundreds of files—all without errors. VS Code even lagged from the sheer speed. In under a minute I got a perfect, zero‑conflict refactor, turning a tedious 20‑minute task into a flash of automation.

I blocked Claude Code’s built‑in Explore agent because it was sluggish, searched the wrong places, used cheap models that hallucinated, and stole context when it read files. By swapping in a custom attention‑agent built on TS‑morph, I got fast, AST‑accurate path results without losing context. The new setup feels far more reliable and speedy than the original.

I tried Claude Code’s built‑in Explore agent and quickly hit its limits – it was sluggish, searched the wrong folders, used a cheap model that spouted irrelevant results, and even caused the main agent to lose context when it read files. After swapping it for a custom attention‑agent built with TS‑morph, I got instant, accurate path lists, kept context intact, and saw a huge speed boost, which made the whole workflow far less frustrating.

I signed up for Claude Team, enabled Memory, and spent hours teaching it about my startup. It even created over 20 memories that I could see in the tool. But after a day and a half, Claude still acts like it knows nothing about me or any reference chats. The feature feels broken, and I’m left confused and frustrated, looking for any fix.

I keep using Claude to draft documentation for work and side projects, but every output is way too wordy and reads like a machine. I constantly have to ask it to “make it simpler” or “use plain language” just to get something that doesn’t scream “AI wrote this.” The whole process feels tedious and the tool’s style is frustratingly robotic.​

I tried using the new Plan subagent, but it kept hiding its analysis and running “in the background.” That meant the agent would wander off for hours, built on invalid assumptions and hallucinations, driving me in the wrong direction. I’m frustrated and just want to know how to turn it off completely.

I tried using Claude to automate API calls, but it kept “mistyping” the access token each time it made a request. Instead of learning from the error and generating a reusable script with variables, it repeated the same typo over and over. The experience was irritating because I expected the model to handle such detail reliably, and the constant slip-ups made the tool feel unreliable for precise tasks.

I tried using Claude Code to fine‑tune UI layouts by sharing screenshots, expecting it to spot exact pixel positions. Instead, it kept guessing, misplacing elements, and only admitted it can’t see pixel‑level details. I had to manually measure everything myself, which felt frustrating and a major limitation for my workflow.

I tried using Claude on Windows, but it kept slipping back to Linux‑style commands and paths. Each time it would “realize” the error, attempt a mount, then fail again with Windows syntax. The constant back‑and‑forth wasted a lot of tokens and my time, and the whole experience felt frustrating and inefficient.

I spent six months pushing Claude Code to rewrite a 300k‑line app, and the tool became my indispensable teammate. I built auto‑activating skills, dev‑docs workflows and PM2 hooks so Claude never lost the plot or left errors behind. The experience was mostly smooth—Claude followed best‑practice guidelines, self‑checked builds, and saved me countless hours—though occasional misfires reminded me to step in and prompt more carefully.

I spent six months hammering Claude Code on a massive rewrite, pushing it to generate 300‑400k lines of TypeScript. After wrestling with missed skill activations and occasional bad outputs, I built auto‑activating hooks, a dev‑docs workflow, PM2 logging, and a suite of custom agents. Those tweaks turned Claude into a reliable co‑pilot—consistent style, zero lingering errors, and huge time savings—though I still have to step in when it stumbles.

I’ve been juggling subscriptions to Claude and ChatGPT for coding, and my experience has been a roller‑coaster. Claude started strong, then sputtered with bad code and hallucinations, so I switched to ChatGPT which initially crushed a big project before it too began repeating mistakes, forgetting context, and inventing variables. After another switch back, Claude is performing well again. The inconsistency leaves me confused and wondering if there’s a deeper reason behind these swings.

I tried using a community‑made subagent for a topic I know well, and it was riddled with mistakes and common myths that Claude usually repeats. The errors made me nervous about trusting any subagents in areas I’m not familiar with. I’d love a curated repo where agents are vetted, tested, and compared to vanilla Claude output, so I can rely on them without second‑guessing.

I spent the day building Duniyar Hausawa, an app for Hausa proverbs, with Claude Code acting like a senior developer on call. It helped me get the core features—2,022 proverbs, search, quizzes, and a clean dark UI—up and running smoothly. The tool’s suggestions made fixing UI quirks and polishing the experience feel effortless, turning a frustrating coding slog into a satisfying progress session.