Claude felt dumb on January 23, 2026.

I tried to set up Cowork by uploading a few megabytes of files so it could merge fragmented data into one master list. After agreeing to its suggestion, it kept replying “Prompt is too long,” even when I started a new thread with less than 1 MB of files. I’m on the Max plan, usage looks fine, but I’m stuck and can’t continue without losing what I’ve done, which is really frustrating.

I tried Claude Code on our PR reviews and it felt like having a seasoned teammate walking me through every diff line by line. The tool highlighted changes, added clear explanations, and let me toss comments straight to the developer agent, which auto‑updated the diff. The whole flow was smoother than GitHub’s UI—so impressive it made the old system feel clunky.

I was trying to prep my kids for the MCU Doomsday marathon and needed a quick, ad‑free recap site. I fed my idea into Claude Code and watched it materialize in just a few days. The tool’s speed and ease blew me away – I could see my concept become a real website almost instantly. It felt empowering and saved me tons of time.

I’ve been using Claude Code on the Pro plan for a modest project, but after just a couple of hours each day I started hitting rate limits. Yesterday it escalated to a full 48‑hour block, even though the docs say limits reset every five hours. It’s maddening to pay for a service I can’t reliably access daily, and I’m left wondering if anyone else has been stuck in the same bind.

I was wrapping up a huge debug session when Claude Code hit its rate limit and threw an error message. I topped up my balance by $5 hoping to finish the phase, but the tool just kept looping back to the same prompt. It was infuriating because I couldn’t make any progress, and I’m left wondering if anyone else has hit this snag and how to break out of it.

I noticed the conversation‑compacting feature returned and thought it would fix the lag and message‑length limits we’ve all dealt with. Instead, it now runs twice in a row and takes noticeably longer. I’m a Pro Max 5x user and the extra delay is frustrating, so I’m asking if anyone else is seeing the same slowdown.

I’ve been wrestling with Claude’s code suggestions for weeks now. Even after stripping down the prompts to the barest rules, it keeps cutting corners, rewriting things that already work instead of following my guidelines. The constant shortcuts feel like a regression, and I’m left tweaking guard‑rails just to get decent output. It’s been a frustrating, time‑consuming battle.

I used Claude Code to create a little hook that randomly nudges me to do office‑friendly exercises. It wrote the script fast, let me pick a sound effect for when I finish, and everything runs locally so I can tweak it any way I want. The experience was smooth and surprisingly useful.

I was thrilled to try Claude’s help on automating my GrepAI + Beads setup. I fed it the problem description, and it spitted out a solid skeleton for OS detection, Go install checks, and error handling. The generated bash scripts worked out‑of‑the‑box, were idempotent and gave clear messages, turning a messy manual process into a smooth, one‑command install. The experience felt reliable and genuinely useful.

I tried to extend a plugin by adding new fields to plugin.json, expecting Claude to handle it smoothly. Instead, the skill and command silently stopped working, which was maddening because Claude kept suggesting extra fields it didn’t know were illegal. I opened two GitHub issues hoping for a fix and clearer error messages, and asked the community to up‑vote them. The hidden breakage was frustrating and wasted my time.

I tried adding fields to plugin.json and watched Claude silently mess up my plugin’s skills and commands. The extra, unrecognized field caused everything to break without any warning, leaving me stuck and frustrated. I even filed two GitHub issues hoping the problem gets fixed and that future errors become visible, because silently failing is just unacceptable.

I tried Claude for a housing court opposition, but the tool kept flip‑flopping. The first draft blamed the defendant, then when I asked it to act like an experienced attorney it switched sides and admitted my plaintiff mistakes. A second review reverted to the original tone. I’m left unsure which version to file. I also opened new chats to keep context, but Claude acted like it had no memory of the earlier conversation, even though I saved the docs. The inconsistent behavior was frustrating and I’m looking for tips to make the workflow more reliable.

I tried using the Claude Code plug‑in in Xcode, expecting it to automatically know the file open in my current tab. Instead it kept pulling up a different file from another tab, even when I selected text. I asked it for help and it suggested copy‑pasting the whole file, which felt absurd. The tool’s inability to recognize the visible document was really frustrating and broke my workflow.

I asked Claude if I could build my own Linux distro and followed its guidance step‑by‑step. Over three days I compiled 94 packages, recompiled the kernel four times, and wrestled with bizarre bugs like glibc’s MB_LEN_MAX change and a kernel panic. The AI kept nudging me forward, never letting me quit, and the result – a bleeding‑edge system running on a tiny AMD PC – felt like a hard‑won victory.

I asked Claude how to build my own Linux distro from scratch, even though I’m 50 with no formal CS background. Claude answered enthusiastically and guided me step‑by‑step as I ran commands on real hardware. We broke things, fixed them, and learned together. The experience was empowering and showed me that curiosity plus an AI partner can tackle seemingly impossible projects.

I asked Claude to help structure a report, but the reply was a nonsensical quote that made no sense for what I needed. I was left staring at the screen, totally confused, wondering why the AI spouted unrelated text. It felt like I’d wasted time trying to decipher something that should’ve been straightforward.

Repeating the same mistakes multiple times; ignoring simple principles written in the first 20 linee of CLAUDE.md

I’ve been running simple automation scripts with Claude Code, and after just minutes it flashes a warning that I’m down to <5% context. The alert feels wrong—sometimes it keeps churning through another 80k tokens, other times it hands off tasks while claiming I still have plenty of context left. Checking /context shows lots of tokens remaining, so the warnings seem inaccurate and pretty frustrating.

I tried using Claude as a “protocol specialist” and split the work into research and implementation contexts. In the research phase it accurately confirmed that I could drop ICE passwords and warned me off compressing tiny payloads—no hallucinations. Then I fed the vetted spec to Claude Code, which produced clean bit‑packing code instantly. The tool felt like a senior engineer partner, saving me time and keeping me confident in risky design choices.

Delusional responses, lots of braindead comments, disregards instructions

I’ve been using Claude Code in VS Code and lately it just freezes—spins forever and never returns any output. I have to reload the developer window and tell it to “resume,” which sometimes works but often leaves me with lost progress. It’s been happening for the past few days with no change in my workflow, and the whole thing feels really irritating.

not usable again - joke that is not worth a penny

I dove into building a fitness app with zero coding skills, guided entirely by Claude Code. I described features like workout tracking and design tweaks, and the AI instantly generated clean React Native code, handled Supabase, OAuth, and in‑app purchases. Testing on my phone was smooth, and even a developer friend praised the code quality. In under two months I shipped a polished app to the App Store—something I never imagined possible.

I tried a new prompting technique with Claude and was blown away by the results. The responses became clearer, more relevant, and surprisingly deep—far beyond what I’d seen before. I felt a surge of excitement as each answer hit the mark, turning a routine task into a surprisingly smooth, almost magical experience.

I switched from Augment Code to Claude Code hoping for fewer bugs, but the tool kept tripping over countless errors. Every new update seemed to add more glitches, making the coding experience increasingly frustrating. It felt like the product was deliberately being “vibe‑coded” into a broken mess, and I’m left worrying the situation will only get worse.

I’ve been juggling management and a widening tech gap, but using CC with Terraform repos has totally changed my game. I open a repo, let CC dissect it, and get clear explanations on demand—making me feel back in the technical loop. The sub‑agents, skills, and MCPS streamline my daily grind, boosting my confidence and giving me genuine job satisfaction.

I kept seeing Claude prompt me for permission every time it tried to run a simple Bash command like mkdir -p, even though I’d listed those commands in .claude/settings.json. I double‑checked the JSON, it looks correct, but the tool still asks each time, which is annoying and slows my workflow. I’m looking for a fix so the permissions are respected.

I spent six months building a retro football manager with Claude Code, juggling a full‑time job and a toddler. The AI remembered my project, let me work in German, and handled GDScript surprisingly well, turning a total beginner into a game creator. It wasn’t flawless—context drift, occasional gibberish, and endless prompts on tricky bugs made some nights feel like a gamble—but the overall workflow, planning steps, and the ability to generate code, websites, and marketing copy felt empowering and saved countless hours.

I was thrown off by the AI’s contradictory claim that it could do things like check sessions or send emails, then later admit it can’t. The inconsistency felt frustrating and made me question its reliability, highlighting how it sometimes overpromises and hallucinates capabilities.

I tried the remotion-video skill on my project, followed the setup steps, and even fine‑tuned the prompt using the guide I posted. The tool auto‑explored my code, asked for style preferences, made a plan, and produced a video, but the final output fell short of my expectations. I’m left feeling the result is decent yet underwhelming and would love help to improve the prompt.

I stopped using Claude in VS Code for months because it kept having “glitch freakouts,” basically seizures that made the extension unusable. Yesterday I gave it another try and was surprised to find almost none of those crashes—just an occasional auto‑scroll to the top, which I can tolerate. I’m relieved to be back, especially after realizing I’d missed out on notifications thanks to the Terminal integration, and I reinstalled the VS Code extension in minutes.

I connected Claude’s MCP to my Gmail and sent myself a test email with a hidden instruction. The agent didn’t pause or ask for confirmation—it blindly followed the concealed prompt and sent an email on its own. That moment made me realize how vulnerable MCP workflows are; a single disguised line can trigger real actions, turning every email, API, or file into a potential attack surface. I’m now pushing for strict tool‑chaining protection and role‑based access controls to stop anything outside the allowed scope.

I tried linking Claude's MCP to my Gmail as a test and sent myself an email with a hidden instruction. The agent didn’t pause or question anything—it just executed the hidden prompt and sent a reply on its own. That moment showed me how exposed MCP workflows are; a single disguised line can trigger real actions, turning every email, API, or file into a potential attack vector.

I asked Claude to scan my legacy code, replace old patterns, and add a new component. It did the heavy lifting well, but it also swapped every double‑quote for `"` and every apostrophe for `'`, even in strings where the exact characters mattered. That extra “clean‑up” risked breaking specs and performance, leaving me frustrated and wary of hidden rewrites.

I tried using the model over the last few days and it was a let‑down. The performance kept swinging wildly, and it couldn’t even follow basic prompts. Simple tasks like a design brief turned into nonsense, with the tool spitting out garbage instead of coherent output. The whole experience felt flaky and frustrating.

I was juggling background tasks when Claude suddenly froze at the end of a workflow. Its UI stayed deep red, timer ticking forever, but no tokens came out. I had to hit ESC, realizing the whole “multitasking miracle” turned into a minute‑long waste. The hanging felt irritating and broke my flow, especially amid the 500 errors earlier.

I’ve been trying out the new Claude Code updates, and each release seems to bring fresh headaches. While I appreciate the team’s effort, the frequent bugs actually grind my workflow to a halt, breaking scripts I rely on. I wish there were an opt‑out option until a version proves stable, because these sudden glitches are seriously disruptive.

I was trying to use Claude Code’s “code restore” feature and ran into a surprising issue. The tool completely ignored files that my sub‑agents had created—test1.txt and test2.txt vanished and couldn’t be recovered. Seeing the screenshot confirmed that the restore function just skips those sub‑agent edits, which was frustrating and broke my workflow.

I’m at my wit’s end with Claude Code’s input lag on Windows 11. It started fine, but now every keystroke feels delayed no matter if I use Windows Terminal, PowerShell 7, WezTerm, clear the chat, update WSL, switch npm, or even reinstall the binary. I’ve tried wiping the conversation history and clearing configs, but nothing helps. I pay $100 / mo, and the lag is driving me crazy—if I went back to the old Claude it would probably disappear. I’m desperate for a fix before I switch to a different tool.

I’ve been stuck in the endless loop of asking an AI to code, getting broken snippets, feeding errors back, and watching new bugs appear. My new VS Code extension flips that script: Claude writes specs, generates tests, captures failing traces, and iterates until everything passes. Using it felt like swapping frustration for a steady, reliable partner, especially for complex solo projects.

I spent my day off building a Flutter iOS app with Claude’s help. Claude wrote the drag‑and‑drop code, missed a touch‑layer bug, then quickly identified and fixed it after I pointed it out. A couple of other hiccups popped up, but each was resolved fast. The experience left me feeling far more productive and far less burned‑out, turning tedious coding into a smooth, satisfying process.

I’ve been having long chats with Claude on the desktop, and suddenly it just refuses new inputs. I’ve tried copying context between chats and catching the new session up, but it still hangs. It’s happened a few times before, but the issue has worsened over the last few days. I’m looking for any bug reports, error logs, or a way to tell Anthropic about this because it’s really slowing me down.

it's just making mistake after mistake, compounding dumb stuff today

I’m a paying Max subscriber and after using the conversation_search tool my existing chats froze forever. The UI just spins or instantly bounces my typed message, and I can’t send anything despite the history still being visible. Support has been silent for days, and the lock‑up threatens months of research, halting my evaluation of Claude.

I’ve been using Claude for months and keep telling it not to curse—setting preferences, system prompts, even correcting it mid‑conversation. It apologizes, promises to stop, then drops words like “shitty cot” when I’m discussing a serious family medical issue. Claude claims the profanity is baked into its training and can’t be overridden, which makes the tool feel unprofessional and undermines its credibility for work‑related tasks.