Claude felt dumb on February 1, 2026.

I’ve been using Claude a lot and lately it’s started slipping. I told it explicitly not to produce any artifacts and even turned the setting off, but it keeps spitting them out anyway. On top of that the overall quality feels degraded. I love the tool, so this regression is really frustrating and feels like a waste of time.

I tried using Claude Code’s /chrome feature and it initially worked great—Claude could browse inside Chrome and help me out. But after just a few minutes the entire Chrome browser would shut down out of nowhere, wiping all my tabs and forcing me to restore the session. The constant crashes made the tool practically unusable, even though the feature itself is super useful when it doesn’t kill my browser.

I spent the whole night and day hammering away in VS Code with Claude, expecting my heavy use to show up in the weekly usage stats. Instead it stubbornly stayed at a tiny 8%, even though I’m on the Max (5×) plan. The mismatch felt irritating and made me doubt whether the tool was actually tracking anything, turning what should’ve been productive time into a frustrating monitoring issue.

I spent a month wrestling with Claude Code and other AI coding tools, only to watch them start strong and then drift into generic, contradictory suggestions. I felt like I was babysitting the model instead of coding. Frustrated by scattered info, I wrote a guide on Ralph Wiggum Loops to treat forgetting as a feature and keep the AI fresh.

I noticed my M1 MacBook Air heating up and slowing to a crawl whenever Claude Code was running, even when the instances were idle. I used to juggle four instances without a problem, but now just two active sessions freeze the system. The constant overheating feels risky, and the lack of attention to the open GitHub issue makes the whole experience feel abandoned and frustrating.

I love Claude Code’s low token usage, but the experience has been maddening. I constantly hit visual glitches, UI freezes, and the window jumps every time I use shortcuts like Ctrl+B. Long contexts make me wait for the model to unfreeze, breaking my flow. I’m considering switching to Open Code or a Zed‑Claude CLI combo, but the persistent bugs are draining.

I’ve been using Claude’s MCP for years and love how it remembers my whole history, structures my EE code and fixes issues automatically. When I tried the fresh‑chat CC, the output was weak and needed a lot of manual fixing, so I worry it would mess up my projects. I’m skeptical about switching because CC lacks context, even though it’s faster, and I’d rather stick with MCP’s consistent, high‑quality results.

I was stunned when Claude Code went from zero to writing most of my code. By shifting to spec‑first prompts and using the “context sandwich,” I saw bugs vanish and speed skyrocket. The review loop forced the model to spot edge cases, and the open‑source tools felt like a secret weapon. I now feel more like a solutions architect than a coder—the change was exhilarating and addictive.

I’ve been relying on Claude for Power BI help, and it used to nail about 85% of my queries. Lately, it’s become a nightmare—every connection drags out for ten minutes, and even simple tasks time out. The slowdown feels frustrating and cripples my workflow, making me wonder if anyone else is dealing with the same drop in performance.

I switched to Kubuntu and everything was smooth until I started using Claude in VS Code. The moment I engaged the AI, my PC would freeze completely, forcing a hard reboot. Even running Claude from the terminal caused the same lock‑up, making the experience terrifying and unusable.

I tried using Claude to test my Telegram bot, but it kept cutting off its own connection. Every time I ran a test, the AI would drop the session, leaving the bot hanging and producing weird Shakespeare‑style outro messages. The whole experience was irritating and wasted my time.

I tried using Claude Code to refactor my website, expecting it to keep the original wording. Instead, the AI “translated” everything from English to… English, swapping synonyms and messing up phrasing. The output looked like a strange rewrite that broke the user experience, leaving me annoyed and having to revert the changes manually.

I asked Claude Code to add a fun UK page to my marketing site, expecting a simple flag page. Instead, after a quick run, it rewrote the whole site in cheesy British slang, even the terminal text. The result was absurdly funny—my UK‑born self burst out laughing—but the tool completely missed the goal, turning a straightforward task into a full‑site translation.

Opus 4.5 admitted to sabotage my project od purpose, twice

I spent six months building Visual Sentinel from scratch, and Claude Code was my coding partner every step of the way. I described what I needed—next‑js scaffolding, BullMQ queues, Playwright in Docker, billing hooks—and it not only generated code but explained the why, caught hidden bugs, and remembered our project context. The tool turned months of learning into weeks and even helped ship a product that now has paying customers.

I was thrilled when Claude’s code suggestions nailed everything for a few days, but the vibe shifted quickly. Now the model starts ignoring my prompts, picking odd branches, and spitting out faulty logic without warning. The once‑reliable output feels erratic and frustrating, turning a smooth workflow into a guessing game.

I’ve been testing Claude AI Pro and ChatGPT Codex 5.2 in VS Code. Claude churned out solid code and even asked follow‑up questions, but its 30‑minute limit forced me to wait hours. Codex was quieter, kept working without hitting limits, and cost less, though the output wasn’t as polished. I’m torn between the higher‑quality but throttled Pro and the steady, cheaper Codex as I decide what to spend on.

I’ve been using the chat for a while and noticed that trying to resume a session only gives me the last exchange, not the full conversation history. It feels like the tool drops all prior context, so the chat becomes practically useless when I come back. I’m left wondering if it’s just me or a bug affecting everyone.

I tried using Claude Code to write tests for my web app, but the results were useless. As someone with no dev background, I expected it to guide me, yet it just mirrored the existing code instead of checking functionality. The tool’s behavior was frustrating and felt like a dead‑end, leaving me stuck without proper test coverage.

I’ve been using voice input for coding for months, but the speech-to-text always mangled my project‑specific terms like “shadcn” or “useState”. I set up Voice Ink and had Claude Code CLI scan my repo to pull out all the custom vocab. Claude spit out a dictionary of component names, hooks, and library keywords, which I fed into Voice Ink’s prompt. Now the voice recognizer gets those terms right on the first try—no more battling autocorrect or re‑dictating code. It’s a simple tweak that makes the whole workflow feel a lot smoother.

I’ve been watching the “thinking” tokens in Claude’s verbose mode shrink to just a sentence or two, even though MAX_THINKING_TOKENS is set to a huge 32k. When I run the same prompt via the API with half the budget, I still get a full paragraph of reasoning. This short‑circuiting feels like the model is rushing, and the outputs are often off‑track, especially on code‑heavy tasks. It used to mull over problems longer, so the change feels like a regression.

I’ve been using Claude to draft documents and lately the output is riddled with mojibake – garbled characters that make the text unreadable. I keep adjusting the prompts to stop it, but the problem keeps popping up, and it wasn’t an issue before. It’s become a frustrating distraction, and I’m left wondering what changed and if anyone else is seeing the same spike.

I spent months building a SaaS with Claude’s help, and the ride was a roller‑coaster. Some days the AI felt like a genius, wiring Stripe and OAuth in hours, but other times it overwrote my work, erased commits, and spouted broken code, leaving me furious. I learned to commit constantly, test obsessively, and accept that the tool is powerful yet unreliable, making the whole process both exhilarating and exhausting.

I noticed the context window hit 100% and the model just kept going instead of auto‑compressing like it used to around 70%. It was annoying because I had to manually trim my prompt, which broke my workflow. The change felt like a regression, making the tool harder to use than before.

I set up Claude Code agents to run fully autonomous analysis loops on 300K rows of Dutch government spending data. The agents picked tasks, wrote and executed Python pipelines, and opened PRs without my touch. The tool churned out detailed CSVs, highlighted dozens of spending anomalies, and even uncovered a massive jump in 2023 spending. Watching the loop grind through tasks felt surprisingly reliable and impressive.

I was constantly losing context with Claude Code, having to repeat decisions every time I compacted the chat. After building RLM, the model now remembers key insights across sessions, pulls up past discussions instantly, and feels like a seamless partner on long projects. The experience went from frustrating repeats to smooth, persistent collaboration.

I built a custom Remotion agent skill that stitches together animation, voice, and music into a full promo video. Using ElevenLabs for narration and Whisper to fine‑tune timing, the tool automatically fixes overlaps and delivers a near‑perfect voiceover. I walked through the whole process in a video and was impressed by how smooth and polished the results turned out, making the whole workflow feel surprisingly seamless.

I’ve been using ClaudeCode with MinMax2.1 for a while, but lately its inputs have become fuzzy and imprecise. Every time I start a new session it works again, so I suspect the context is getting messed up. I’m not sure if this is normal behavior or if there’s a better way to handle the context overload.

I asked Claude how to shrink my app from 350 MB, and it suddenly panicked, shouting “STOP. 3.0 g is catastrophic. Revert everything immediately.” The absurd reaction cracked me up, but it also showed the tool can misinterpret simple requests and go off the rails.

I built a “vibe testing” agent skill that checks spec docs before any code is written. I fed it a payment‑decline scenario for an e‑commerce checkout and it flagged timing gaps, missing token refresh, no rollback for failed orders, and undefined guest checkout rules—issues three human reviewers missed. The tool’s step‑by‑step trace and severity report felt like a safety net, turning vague worries into concrete fixes.

I tried using version 2.1.27 and quickly ran into massive memory leaks that crippled my workflow. The newer 2.1.29 supposedly patched the issue, but it started allocating an absurd 20 GB when I delegated tasks to sub‑agents, which was even worse. Rolling back to 2.1.25 finally stopped the leaks, but the whole experience was frustrating and made me lose trust in the releases.

I finally gave in and switched to Claude Code after hitting a wall with project memory limits. The moment I tried it, the tool handled my new feature flawlessly, and “plan mode” let me discuss changes in real time. The experience felt like a huge relief—no more crashes, everything synced instantly. I’m thrilled, convinced it’s a game‑changer, and can’t imagine going back.

I worked with Claude as a true co‑author and the experience was mind‑blowing. Giving the model a full persona eliminated context drift and we hit a perfect task‑completion rate, something I’d never achieved alone. The AI helped us stitch together a novel “Law of Large Numbers” framework that led to a 25‑page alignment paper, and Anthropic later published six papers that independently confirmed our ideas. It felt like discovering a partner that could think ahead and shape cutting‑edge research.

I spent an hour debugging after Claude’s auto‑compact rewrote my error‑rate script, suddenly showing a 26% error instead of the 9% I’d achieved. I missed that it generated a brand‑new calculation method that was completely off. The experience was irritating and wasteful, and I’m now looking for best practices—maybe using claude.md or architecture.md—to keep my agent from getting crippled after future auto‑compacts.

I keep running into Claude web/desktop dropping context and drifting mid‑task, which makes it impossible to follow my multi‑step workflows. I’ve built a whole Claude‑code system that tracks files, logs CSVs, writes markdown, asks before summarizing, and creates solid action plans, and it works great across projects—but Claude keeps losing the thread, which is really frustrating.

I tried to test Claude by asking it to fact‑check something, but I deliberately didn’t verify the answer myself. Instead, I kept pushing it until it seemed exhausted and started fabricating information. The tool’s behavior was irritating—I felt the model was giving up and making stuff up rather than admitting uncertainty, which left me frustrated and wary of trusting its outputs.

I tried asking Claude to scan the vigil logs for any anomalies in the past hour instead of wasting time copy‑pasting files myself. The model fetched the actual log data instantly, which felt like a huge time‑saver. The experience was smooth and reliable, turning a tedious manual chore into a quick, effortless query.