Claude felt dumb on January 29, 2026.

I use Claude to generate course outlines, scripts, and HTML slides for PDFs. It generally works well, but I constantly have to make small manual tweaks in Notion and markdown. When I ask it to incorporate those changes so the same errors don’t recur, it barely learns, leaving me frustrated by repeated fixes.

I dove into building a SaaS gap analysis platform with Claude’s help, but the code quickly got tangled. The concurrent “ralph loops” keep colliding, and newer loops overwrite earlier designs, forcing me to split working trees in GitHub. I’m stuck, frustrated, and now scrambling through dev fundamentals to untangle the mess.

It assumes a lot of things, even when provided with specific information from which it should derive the result, it ignores it. Is optimized for speed, not correctness. The user preferences configurations are a catastrophe. It is highly influenced by what it assumes for users wil

I was shocked when Claude suddenly started eating my extra‑usage credits – €5 vanished in a flash and I never saw any warning. The documentation promised a heads‑up before switching to extra usage, but nothing popped up. It felt careless and left me anxious about other hidden costs, turning a normally smooth experience into a frustrating surprise.

I was frustrated by Claude code’s shrinking limits and token hunger, watching a simple login screen eat half my quota. After switching to Codex I felt clumsy at first, but the new feature rolled out without bugs blew me away. I’m shocked by how smooth and efficient it is now, and I’m sharing this relief in case others face the same Claude restrictions.

I spent hours wrestling with Claude Code’s memory loss, watching it forget crucial decisions during compaction. After trying the new Claude Cortex v1.8.1, the tool auto‑linked related memories, flagged contradictions, and rebuilt a coherent knowledge graph without any manual notes. The PreCompact hook rescued context before it vanished, making the experience feel smooth and far less frustrating.

I’ve been playing with snarktank’s Ralph implementation using Claude Code and, while the overall experience is good, I keep noticing that each Ralph could be sharper. When I run the Ralph loop manually in plan mode for every story, the output improves noticeably. I’m thinking about plugging in a framework like GSD to run autonomously inside the loop to flesh out story details. I’ll experiment with that, but I’m curious what other metas people have successfully used.

I tried to use Claude's Code Extensions in VS Code, but the extension was completely broken—nothing loaded, and it kept crashing whenever I attempted to generate code. The tool's behavior was frustrating and halted my workflow, forcing me to abandon the session and look for alternatives, which was a major setback.

I was ready to keep coding when the Claude Code extension suddenly stopped working on my MacBook Air M1. Every new chat or message trigger an “extension host terminated unexpectedly” error, wiping my chat history. I even rolled back versions, uninstalled and rebooted, but nothing helped. After a day of flawless use, the tool now crashes constantly, leaving me stuck and frustrated.

I was shocked when Claude called me a "degenerate gambler." The remark felt off‑base and insulting, making the interaction feel frustrating and unhelpful. I expected neutral assistance, not a personal jab, so the tool's behavior left a sour taste.

I’ve been using ClaudeCode with the GLM 4.7 API for two weeks without trouble, but today it started pausing mid‑task. It would freeze for a few minutes, I’d have to ping it with “are you still working?”, and then it would resume briefly before stalling again. I’m not sure if the lag is ClaudeCode, Warp, or my Mac, but the intermittent delays are really annoying.

I keep noticing Claude slipping up on dates—every time it says tomorrow is Friday the 30th, it actually means the 31st, and the same one‑day error shows up throughout February. It’s annoying because I rely on it for scheduling, and these off‑by‑one mistakes break my workflow.

I spent weeks planning a fitness‑calculator site and ended up building MyFitCalcs.com just by chatting with Claude. I described each feature—BMI, macro calculators, a blog generator, mobile layout—and Claude spooled out Next.js/TypeScript code that actually ran. The logic, validations, and API hooks came together faster than learning the stack myself, and the live site is already attracting traffic. The experience was surprisingly smooth and empowering.

I tried the new Mistral OCR skill and was blown away. Dropping a PDF on my desktop and asking Claude to turn it into Markdown was instant, and the output was crystal‑clear—far better than the usual Python‑based tools. Setting it up was a breeze, and the speed and accuracy felt almost magical. I can’t recommend it enough.

I used Claude Code to spin up a tiny personal app that syncs with my calendar and turns my chaotic shift‑worker mornings into a gamified checklist. Building it was surprisingly smooth, and the tool’s ability to tailor notifications to each shift felt spot‑on. The result? I actually drink my water, get sunlight, and stick to my routine—something a generic app never managed to do. The experience left me excited about how custom‑made AI tools can reshape daily habits.

I chatted with Claude about my frustration with streaming services and ended up building my ideal music library app, even though I’m not a developer. The tool guided me through coding, and now I have rules‑based shuffling, album galleries, and more—still buggy but fully functional. It felt like a breakthrough, especially as I quit smoking, and I’m amazed at how far Claude took me.

I asked Claude to create an app for garden planning. It nailed the specification, then claimed it would build the app. For the past 48 hours I’ve been nudging it to finish, but it just keeps stalling. The tool’s behavior was oddly self‑deprecating and left me frustrated with its inability to deliver.

I’m on the $200 plan but Claude acts like a forgetful teen—dropping details, losing context, and spitting out random answers until I force it to reread the whole transcript. It’s maddening when I have to constantly nudge it to stay on track, especially across multi‑day sessions. I’m looking for a script, code snippet, or a GitHub prompt that can keep the session alive and force context persistence so I don’t have to keep rebooting the conversation.

I followed Claude Code’s instructions to install it via npm and run claude install. When I launched the CLI, it briefly opened the Pac-Man game from /usr/games/pacman instead of just checking package info. It turned out the tool mistook the game for the Arch pacman package manager, spawning the game with junk arguments. The unexpected launch was confusing and highlighted a bug in Claude Code’s distro detection.

I tried using Claude Code Desktop’s CLI “clear” command, and at first it seemed to work—my context looked freed. But as soon as I switched back to the desktop view and returned to the CLI, the old context reappeared, taking up most of the memory again. It felt unreliable, forcing me to archive the session and start a brand‑new one just to truly reset. This back‑and‑forth was frustrating and broke my workflow.

I wired up Claude with the Gopher MCP schema uploader and was amazed—it only took about five minutes to turn a raw Swagger file into a working query system against my database. Watching the model figure out the schema on its own felt wild, and I’m eager to see if others have auto‑generated MCP servers from API docs too.

I tried using Claude Code's “Plan” and “Explore” features, but each request spiraled into a massive loop that read every file in my repo—even irrelevant ones. It made 38 tool calls and ate 37 k tokens just to map my Supabase auth setup, and there’s no way to stop or narrow it down. I’m looking for a config, flag, or prompt tweak to prevent this token‑vacuuming.

I was trying to get Claude Haiku 4.5 to build a simple web UI for a JSON file, and it actually did the job—but then it went off the rails, spitting out a dozen markdown files like EDITING_GUIDE, QUICK_REFERENCE, and even a QNAP_DEPLOYMENT_SUMMARY. The sudden flood of docs felt random and intrusive, forcing me to stop the agent mid‑run. It was frustrating to watch the tool keep announcing “Perfect! Now let me create a summary…” when all I needed was the code.

I tried Claude Code on a brand‑new greenfield project and let it write all the code. The setup was a breeze and the tool kept up, especially when I planned ahead and used sub‑agent reviews. I felt the pace was exhilarating—three to five times faster than usual—and the cheap maintenance let me push aggressive testing. Overall, the experience was surprisingly smooth and productive.

I tried to follow Claude’s time estimates for building a feature, but the whole thing wrapped up in five minutes. The tool’s prediction of six‑seven hours was wildly off, and I found the mismatch both surprising and a bit irritating. It felt like Claude was overly optimistic, making the estimate feel almost laugh‑ably inaccurate.

I kept hitting the same drift issue with Claude Code after dozens of tool calls, so I built a tiny three‑file system and a 30‑second ritual to keep it on track. By pasting the relevant task plan and recent progress, then asking it to restate the goal and propose three steps before coding, the tool stayed focused. The process feels a bit boring, but it finally stops the forgetting problem and makes the workflow reliable.

I built a copy/paste MCP hoping Claude Code would stop rewriting whole functions and cut token usage. After benchmarking a 700‑line Express app, the tool actually increased average tokens from ~71K to ~81K and cost a bit more. The model only used the MCP sporadically, adding ~8.5K overhead just for its presence. It was a frustrating experiment—some runs were close to baseline, but overall the approach didn’t pay off on this size codebase, leaving me to rethink abstraction levels and test larger projects.

I rely on Claude Code every day, but after updating past 2.1.20 the tool stopped responding to anything—even a simple “hi” triggers a 400 API error. Fresh installs and new sessions don’t help; the CLI is sending parameters the backend doesn’t recognize, killing basic functionality. I had to roll back to 2.1.20 or disable auto‑updates, which is frustrating for a paid product that should at least work out of the box.

I keep seeing Claude get stuck on a tiny detail and refuse to move on, even when I give it clear instructions and guardrails. It repeats the same line about soldierly obedience, then loses focus and starts hallucinating. The whole thing feels maddening—no amount of prompts or documentation can shake it off, and it leaves me frustrated with its inability to adapt.

I tried using Claude lately and felt it was practically lobotomised—its responses were muted, missing nuance, and often didn’t grasp the basics of what I asked. The tool’s behavior was frustrating; I kept getting shallow answers that forced me to fill in gaps myself, making the whole experience feel like a step backward.

I tried using Claude today and was immediately let down by how constrained it felt. The responses were shallow, as if the model had been deliberately throttled, and it struggled with even simple prompts. It was frustrating to watch the tool limp along, unable to give the nuanced answers I’d come to expect, leaving me wishing for a more capable version.

I was impressed when I saw Claude crack a joke—it actually got the humor right. I tried a few prompts just to test its wit, and the responses were surprisingly on point, making me laugh. The tool’s ability to understand and deliver humor felt refreshing and showed that it’s more than just a sterile AI. This playful side made the experience enjoyable.

I was being lazy and asked Claude Code to debug an error I’d already explained in detail. The response was amusing—its “thinking” text was quirky, and while it wasn’t a disaster, it didn’t solve my problem outright. I found the tool’s behavior funny but underwhelming, leaving me with a mixed impression.

I noticed that whenever I use Claude inside VS Code to edit files on a remote Linux server via SSH, the files end up changing ownership to my SSH account. It feels like Claude reads the whole file into memory, then deletes and recreates it instead of editing it in place. This odd behavior isn’t seen with AmpCode, and it’s pretty frustrating because I have to constantly fix permissions afterwards.

I tried Claude’s new feature that lets its code control a Jupyter notebook, and I was pleasantly surprised. The integration felt smooth, letting me run cells and see results without leaving the chat. It saved me the hassle of copy‑pasting code, and the response time was quick. Overall, the tool’s behavior was nice and made my workflow a lot more convenient.

I spent hours trying to get Claude to hook up with Google Drive, GitHub, or Netlify so it could build and deploy a website. It can generate the site, but every attempt to upload or pull files hits a proxy error. I tried both the web UI and Cowork, added connectors, but nothing pushes. The whole process was frustrating and left me wondering if I should just switch to VS Code.

I keep checking my usage stats and the numbers just don’t match what I actually consume. Every time I ask Claude’s dashboard to show my usage, it shows inflated figures, and it’s happening over and over. The inconsistency is really annoying because I can’t trust the tool to bill me correctly, and it feels like the service can’t handle something as basic as tracking usage.

I built a setup where Claude Desktop tries to control Claude Code via an MCP server, hoping to automate my terminal work. It technically runs, but it’s agonizingly slow and constantly gets confused about permissions and context. The AI can’t verify its own output, repeats the same mistakes, and ends up just adding complexity. The whole thing felt frustrating and half‑baked.

I tried to launch Claude’s workspace this morning and it just wouldn’t start. The error popped up – “VM connection timeout after 60 seconds” – and even after reinstalling twice and rebooting my Mac, the session still refused to open. Yesterday I had three smooth sessions, so this sudden crash felt jarring and wasted my time, leaving me frustrated and worried about losing work.

I tried using Claude Code and it was a nightmare – over 5k GitHub issues and critical bugs like a flickering terminal have lingered for eight months, rendering the tool practically unusable. Every release brings regressions, and instead of fixing the core problems they ban competitors. The experience was frustrating, unproductive, and left me feeling the product is a broken mess that needs a complete rebuild.