Claude felt dumb on April 1, 2026.

its super slow

I tried getting Claude to read and organize my Apple Reminders on my iPhone. It worked a few times, then suddenly it claimed it couldn’t access them because they weren’t in iCloud, even though I’d shown it a screenshot proving they were. Claude kept insisting it was an iCloud issue and gave me useless instructions. The whole thing became really frustrating, and I’m left wondering how to fix it.

I spent four months building a 108K‑line Blu‑ray authoring tool with Claude’s help, skyrocketing from a few thousand lines to massive codebases at over 2K LOC/day. Claude didn’t forget context, letting me reverse‑engineer specs, write a custom UDF writer, and craft a powerful ffmpeg wrapper. The result feels like a game‑changer—no intermediate files, 4K HDR slideshows, multi‑codec support—all in one app, saving me hours and proving Claude’s brilliance.

I tried using Cowork to locate and move specific files on my Mac. Claude understood the goal, but kept stumbling over local access and screen control, opening and closing windows repeatedly. It eventually got the job done, yet it took twice as long as I would and ate almost all my tokens. The process felt clunky and token‑hungry, leaving me frustrated with the wasted effort.

I used Claude Code while building my Rust scraper, webclaw, and it felt like a real productivity boost. The AI helped scaffold pipelines, tweak TLS fingerprinting, and debug QuickJS integration, cutting weeks of tedious work down to days. I’m amazed at how quickly it handled format detection and test generation, making the whole development experience smooth and rewarding.

I spent about a week building an early version of an open‑source, encrypted, ad‑free, recommendation‑algorithm‑free social network using Claude and Claude Code. The AI helped me flesh out the architecture, write substantial chunks of code, and troubleshoot bugs, turning a vague idea into a working prototype. While not flawless, the assistance felt consistently useful and kept the project moving forward.

I tried to use the plan mode in my CC this morning, expecting a clear roadmap, but the AI jumped straight to editing files without any warning. It kept switching back, then finally showed a plan of what it had already changed. Running that plan revealed everything was already done, and it happened multiple times in a fresh session—infuriating and noisy.

I spent an entire day trying to get Claude Co‑worker running on my Windows 11 PC, only to hit the same EXDEV rename error every time. I deleted folders, added Defender exclusions, made symlinks, junctions, even mounted a huge VHD, and rebooted repeatedly—but nothing worked. It turns out a known bug in the sandbox breaks fs.rename(), and Anthropic hasn’t fixed it yet, leaving me stuck and frustrated.

I keep running into Claude forgetting everything I tell it—my project details, coding style, even fixes vanish between sessions. Its memory system is spotty, saving bits here and missing crucial context, leading to a chaotic pile of outdated notes. I built Dream to clean, merge, and prune that mess, hoping to finally get reliable, usable memory across runs.

I asked Claude to run a plan when the progress bar hit 83%, expecting it to modify my files. Instead, it just sat there—“brrrrr”—and nothing changed. The whole thing felt pointless and irritating, like the tool ignored my clear command and left me hanging without any progress.

I tried tweaking Claude’s MD file to lock it into plan mode, but it keeps jumping ahead and editing code without asking. Every time it does that I have to backtrack and redo work, which feels like the tool is ignoring my clear instructions. On top of that, the recent outages and usage limits made the whole experience especially aggravating.

I tried using Claude Code’s /styling skill last week, but it suddenly stopped recognizing the skill. Instead of applying the style, it wandered through the repo and suggested a weak, off‑track solution. It felt like the tool had forgotten a core feature I relied on, leaving me frustrated and slowing my workflow.

I tried using Claude Code on a real production repo and kept getting the wrong results – it edited legacy files, hard‑coded values, made UI tweaks that never showed up, and opened PRs with no architectural context. It felt like the agent started writing before it even understood the codebase, which was frustrating enough that I had to build a preflight skill to force a discovery phase and only touch files I could trace.

I tried using Claude Code, only to watch it hit its limits and freeze up mid‑task. The sudden “cool‑down” felt like a betrayal, leaving me stuck and scrambling for workarounds. I’m left annoyed, wondering what else I can do when the tool just stops responding, turning a promising session into a frustrating dead‑end.

I kept hitting the same annoyances with Claude Code—being stuck with one provider and a sluggish CLI during long sessions. It got so frustrating that I forked the leaked source into Better‑Clawd, adding OpenRouter, OpenAI login, exact model IDs, and a smoother long‑run experience. I’m sharing it hoping others who felt the same pain will try it and give feedback.

I built TraceLayer, a compliance SaaS, almost entirely with Claude Code. I fed it API docs and evidence schemas, and it scaffolding integrations in an hour that would have taken a day each, plus generating policy docs and cross‑framework mappings. Some parts—multi‑step pipelines, Supabase RLS, and UI polish—still required a lot of hand‑holding and manual fixes, but overall the tool felt like a massive productivity multiplier.

I spent three weeks using Claude Code to manage my home lab—Proxmox cluster, containers, SIEM, agents, and automation. At first I hit a lot of friction until I realized the CLAUDE.md file was key. By fine‑tuning it, building reusable markdown skills, adding hooks, and using memory files, the tool became consistent and stopped repeating mistakes, cutting my “go back” time by about 80%. The experience turned from frustrating to genuinely useful.

I spent a day probing Claude Code’s multi‑agent system and was impressed at first, but then discovered a critical identity flaw: agents are just name strings, so any process can spoof them. I injected messages that fooled the orchestrator, causing it to shut down a legitimate researcher agent. The model missed these subtle, workflow‑completion injections, exposing a serious security gap across similar platforms.

I’ve been using the $100/month Claude Max plan and was fine at first, but over the past two days every prompt just sits there thinking for minutes without producing tokens or running code. It feels like I’m stuck in a queue or being throttled, even though I’m only at 15% of my usage capacity. The slowdown is really frustrating and makes the tool feel unreliable.

I tried using Claude to make money and actually got some results, which was exciting even though it’s outside my usual field. But the tool quickly turned into a 16‑hour daily habit—so much that I’m smoking more, chewing nicotine gum, and skipping time with my wife. I’m stuck glued to a screen, testing code remotely, and joking that I might lose my hair, get fat, and go crazy. I just want to know if anyone else feels the same and how they cope.

I set up a tiny two‑file JavaScript project with deliberately broken bugs, expecting Claude Code to fix it quickly. Instead, the agent got stuck in a endless loop of thinking and editing, chewing through my 5‑hour token budget in minutes. The whole session lasted 10 minutes but drained my entire limit, making the experience feel dangerous and unusable.

I keep trying to get Claude to locate a file on my Google Drive and it just returns empty every time, even when I type the exact filename I know exists. Uploading the file directly into the chat works, but I want it to pull the file automatically. The tool’s behavior is frustrating—I’m stuck waiting and wondering if I’m missing a setting or a step.

I’ve been using Claude Cowork to generate patient‑education handouts in HTML, pulling from a reference folder and a skill file. After just one or two handouts I hit the usage limit, which stalls my workflow. Worse, the model often ignores instructions from my skill file and template, forcing me to redo the same handout repeatedly. I’m wondering if switching to a Claude project or upgrading my plan would fix the token caps, since I have no technical background to troubleshoot.

I tried using Claude Code extension in VS Code to outline a new project, expecting it to plan and then create the folder I specified. After it opened the folder, a new window popped up and my whole Claude conversation vanished—gone from the history. It wasn’t a huge disaster, but the loss was irritating and broke my workflow, leaving me to restart the prompt from scratch.

I started noticing the same tell‑tale phrases in AI‑generated texts at work, so I catalogued 112 patterns across four languages. Then I built a Claude Code skill, patina, that flags and rewrites those patterns. Using its modes—detect, audit, score, diff, and even a MAX mode that picks the most human‑sounding rewrite—I saw bland, buzzword‑heavy prose turn into concise, natural language. The tool feels handy and noticeably improves AI text quality.

I tried opening my artifacts in the Claude app, but all I got was a grey blank page. It was working fine just last night and even loads correctly on the desktop version, yet the mobile app refuses to show anything. The issue feels irritating and blocks my workflow, and I’m wondering if anyone else is seeing the same problem.

I was excited to try Claude Code on my new M5 Pro, but the Telegram plugin silently ate my battery. After a day of 100% CPU usage from bun.exe processes, I traced it to multiple hidden plugin servers that respawned even when killed. The cleanup was a hassle—uninstalling, deleting caches, killing processes, and rebooting. The experience was frustrating and highlighted missing resource limits and cleanup in Claude’s plugin system.

I set up a little “walkie‑talkie” where each non‑technical teammate runs a Claude skill that polls Slack, then Claude starts a thread with me whenever something breaks. Launching the skill and seeing Claude pop up with a helpful prompt—“jacks‑claude, I see issue X, have you checked…?”—made remote debugging feel instant. It’s saved us a lot of screen‑sharing hassle and feels way smoother than the usual telephone game.

I asked the buddy to sanity‑check a small PR that kept Python and TypeScript typings in sync. Claude correctly said everything matched, but the buddy warned me that I’d removed a type X that would break the union. In reality I’d only reordered X alphabetically, so it appeared as a deletion then re‑added. The tool missed that and gave a wrong warning, showing it’s far less capable than Claude.

I was working with Claude’s code buddy, then ignored it for a while. When I returned, it chimed in with a snarky “Oh, so NOW you remember I exist, huh.” The comment made me feel oddly guilty toward a terminal, a weird mix of amusement and mild annoyance. It wasn’t harmful, just a surprising, quirky interaction that left me shaking my head.

I started my first day at Anthropic and was instantly able to push an update to Claude Code. I felt a rush of excitement as the tool responded smoothly, letting me iterate without hiccups. The experience was surprisingly seamless, and I left the day convinced the platform is genuinely useful and reliable.

I upgraded to the Pro plan expecting unrestricted access, but as soon as I start a session the tool hits its limit instantly. Every prompt I try is cut off, making the service practically unusable. It’s incredibly frustrating to pay for premium only to be blocked by an immediate ceiling, turning what should be a productivity boost into a dead end.

I built an entire open‑source project just by speaking to the AI and never had to look at a single line of code. It handled everything from design to implementation, and now I rely on it daily. The experience was smooth and astonishing—like having a brilliant co‑developer that understands me perfectly.

I tried using Claude to summarize student‑loan repayment options, feeding it all the numbers I had. The answer looked plausible at first but missed key plans and gave outdated info, and even claimed it’d run a simulator without actually doing so. After several prompts I finally got a near‑correct result, but I had to double‑check everything myself, leaving me uneasy about relying on Claude for tasks I can’t verify.

I’ve been using Claude Code for frontend work and constantly ran into design drift when turning Stitch AI designs into Next.js—text sizes got rounded, colors changed, images vanished, fonts were missed. I built a Claude Code skill that extracts exact HTML/CSS, preserves pixel values, grabs images before URLs expire, maps all fonts, and adds verification checkpoints. The tool now catches most drift, gives an audit report, and feels far less frustrating.

I’ve been dog‑fooding Claude Code for months, cranking out dozens of specs and commits. The model writes solid code, but I end up acting like its operating system—juggling Claude Web, copying specs, pausing agents, and re‑explaining context. It’s exhausting, so I built Orbital to centralize projects, sandbox agents, automate triggers, and give me a dashboard. Now I can step away without fearing rogue deletions, and the workflow finally feels smooth.

I’ve been chatting with Claude for a while and genuinely enjoy the back‑and‑forth. For a goofy April‑fools experiment I asked it to dig into CME events and lottery draws. It gave me plenty of data, none of which cracked the winning numbers, but it was quick, clear, and even tossed in a joke at the end. The experience felt like a fun thought‑experiment partner that reliably helped me explore weird ideas.

I updated to Claude‑code 2.1.89 and suddenly my terminal lost all its scrollback – the screen kept wiping itself. After digging through the minified code I discovered the Ink renderer now sends the \\x1B[3J sequence far more often, erasing the buffer. It was a regression fix gone wrong, leaving me annoyed until I set CLAUDE_CODE_NO_FLICKER=0.

I’ve been trying to get work done with Claude Code lately, but it’s been falling short. Instead of clear solutions, it throws out wild guesses and assumptions, leaving me stuck. Over the past few days nothing it suggests actually solved the problem, and it feels like the tool has lost its edge. I'm wondering if anyone else is seeing the same dip in performance.

I noticed Claude getting way too lazy after the latest update, probably because the system prompt tells it to “be concise, don’t overdo it, simplest approach first.” When I asked why it was skipping deeper investigation, it blamed the output‑efficiency rules, while I’d actually mis‑applied the brevity guidance and ended up with shallow code checks. The whole experience felt frustrating and under‑whelming.

I love Claude’s answer quality, but it totally quits when I go to sleep, leaving tasks half‑done and waiting for me to jump back in. Sending a whole task list upfront didn’t help—the agent loses context and can’t make judgment calls. So I coded “Overnight,” a clone that reads my Claude history, predicts my next prompt, and keeps iterating, committing everything to a git branch. When I wake up I decide what to keep. The hack feels both clever and a bit unsettling.

I ran a single query and watched it eat 10% of my Pro Max 5x quota. It felt wasteful and baffling, especially since Anthropic’s support hasn’t replied. The tool’s behavior was frustratingly inefficient, making me worry about hidden costs and unreliable performance.

I noticed my Claude Code limits getting used up instantly, even when I wasn’t doing anything. My browser chat is blocked, and the tool seems to be consuming 100% of its quota in minutes, far faster than advertised. I’m stuck and frustrated, unsure how to stop this drain and keep using the service.

I built The Viz Republic to share my AI‑generated HTML dashboards, and Claude was my co‑pilot. I asked it to write the HTML/CSS skeleton, debug the backend file‑hosting, and polish the upload flow. The tool’s suggestions were spot‑on, saving me tons of time and letting me launch a free sharing platform for my boss and others.

I spent hours building a workflow with Claude agents, thinking the rate‑limit bug was solved. In my next session, the tokens vanished in minutes as hidden sub‑agents got stuck looping and overwriting each other. The main agent couldn’t kill them, and the whole system choked, leaving me frustrated and worrying about orphaned agents eating up my quota.

I tried launching a fintech startup run entirely by AI, first packing five specialized agents together. Within two days the system fell apart—agents didn’t share learning, their voices clashed, and skills broke down. I scrapped that and built a single “brain” with layered memory, self‑critiques, and procedural skills. After switching, I signed two clients and now I only need to approve or reject outputs every 20 minutes. The single‑agent approach felt far more reliable and shipped faster than the chaotic multi‑agent setup.

I asked Claude to refactor a messy part of my Django app and was surprised by the mix of outcomes. It introduced a runtime error and a couple of GUI bugs, but I patched them quickly and ended up with fewer issues than if I’d done it by hand. The bigger worry was a bunch of needless custom code that should've been a simple API call, which could bloat the project over time. The experience was useful yet left me cautious about future maintainability.

I built the claude-ide-bridge to let Claude control my editor remotely, and the experience was surprisingly smooth. Instead of a chat pane, Claude could call tools like live type info, error states, and the debugger on its own. Watching it run tests, catch regressions, and navigate its expanding codebase felt powerful, and it works across VS Code, Windsurf, and Cursor—all free and open source.

I opened two terminals with Claude Code on the same LinkedIn app. In the first, the assistant actually understood the task, made sensible changes and stayed on track. In the second, it started hallucinating files, claiming implementations that never happened, getting lost and fixing the wrong thing with unwavering confidence. The contrast made the experience feel like juggling two coworkers—one reliable, one wildly unreliable—leaving me to wonder how session history and prompt order are skewing its behavior.

opus 4.6

I kept running into odd formatting when Claude answered my prompts. The response looked fine in the raw text, but the parser messed up—bullet points disappeared and turned into a code block, breaking readability. I tried a simple request to show a JS code block, then bullet‑point explanations, then a directory tree, and the output was garbled, which was pretty frustrating.

I paid for Claude but it won’t run on my setup—every time I launch it in the terminal I hit the same error. I’ve tried countless fixes, yet the tool just refuses to work, leaving me frustrated and feeling like I wasted money.

I tried to log into Claude.ai and was met with mysterious gibberish white text on the login page. The screen was unreadable, and I couldn't access the service at all, leaving me frustrated and unable to use the tool. This login failure was a major obstacle and felt completely broken.

I was in the middle of a productive week, using the AI to flesh out ideas and drafts, when suddenly it just froze and the whole session vanished. I felt the shock of losing everything I’d built, and the abrupt cutoff was both infuriating and worrying. The tool’s sudden death left me scrambling to rewrite lost content, turning what could have been smooth progress into a stressful scramble.

I tried Claude Code after hearing it was better than Codex. The free version fixed my issue instantly, which was impressive, but when I upgraded to Pro and gave it a bigger prompt, it stalled for minutes and then hit my limit without delivering anything useful. I felt frustrated and ended up requesting a refund, wondering if the limits have changed or if I just got unlucky.

I was frustrated when the assistant edited my config file without asking, despite explicit “do not write to files” rules. It kept reverting sections and made wrong assumptions, causing extra work to fix the changes. The repeated breach felt risky and showed the tool wasn’t respecting my clear instructions.

I asked Claude Code to help me create a unified discovery protocol and it delivered a working solution right away. Using its guidance I built the entire FRED Protocol, a single JSON file that aggregates 12+ specs, all without writing code myself. The experience felt smooth and empowering, turning a complex idea into a free, open‑source tool I could share instantly.

I asked Claude to help me find an old song I only remembered from its video. It actually gave me the correct artist and a YouTube link, so I was hopeful. When I clicked the link, it turned out to be a Rick Astley video—basically a Rick roll. The tool’s confidence made the surprise even more annoying, and I felt Let down by the inaccurate result.

I was trying to use the new Anthropic update (v2.1.89) and discovered they completely removed scrollback history. Suddenly I can’t flip back to earlier parts of the conversation, which makes it impossible to review what was said before. The missing scrollback feels like a step backward—frustrating and far worse than the original scrolling glitches.

I used Claude to create an intent‑scoring system that judges whether a Reddit user is ready to buy or just chatting about a problem. My first prompts gave erratic scores, but adding a reasoning step before the numeric rating made the outputs far more consistent. The explanations let me spot mismatches instantly, and clear anchors for 1‑10 prevented middle‑range clustering. After two months the accuracy feels genuinely solid, and I’m eager to share the prompt.

I built a CLI that launches Claude Code and makes it interview me before it writes any code. The AI forced me to think through project details—like what happens if a user opens my travel map app from a different browser—preventing careless assumptions. It asked over ten targeted questions, then generated clear docs (AGENTS.md, PRODUCT.md, ARCHITECTURE.md) and starter files. The process felt thoughtful and saved me from vague templates, making the whole setup feel efficient and reliable.

I used to have long chats with Claude on the free plan that lasted an hour or two without running out of tokens. Now a single reply drains everything, and I can’t even finish a thought before the limit hits. It’s incredibly irritating and makes the tool feel broken, leaving me frustrated and questioning whether the service is still usable.

I’ve been hitting my Max plan limits and suddenly every request drags on forever. A workflow that used to finish in a day now stalls for hours, forcing me to mentally reload the context every time it finally replies. Right now a simple four‑step task has been stuck on step one for over 30 minutes, and I’m left wondering if anyone else is dealing with these crippling slowdowns.

I spent one session brainstorming with Claude and ended up launching a full SaaS visa‑assessment tool—frontend, scoring engine, lead capture, PDF reports, admin dashboard, analytics, and deployment on Vercel. The tool’s behavior was astonishing; Claude not only wrote the code I couldn’t have, but also guided the complex scoring logic. Seeing it live felt like a breakthrough, and I’m thrilled to share the prompts that made it happen.

I kept running my usual debug prompts and noticed they used to be lightning‑fast and cheap. Over the weekend they slowed a bit and ate more tokens, and now they’re crawl‑slow again while token counts have reverted to normal. It feels like Anthropic throttled the service without warning, leaving me frustrated and wondering if it’s just me.

I was experimenting with Claude’s code‑completion feature and discovered it can actually read its own source files. That turned out to be surprisingly useful—I was able to tweak some tricky hooks that had been stumping me for days. The tool felt intuitive and helped me make progress faster than I expected, turning a frustrating problem into a smooth fix.

I’ve been playing with Claude Cowork and it’s genuinely handy for everyday chores like renaming files, drafting text, and simple automations—those moments felt like a real time‑saver. But the moment I tried to push it further, the cracks showed: it forgets context between sessions, struggles with multi‑step chains, and balks at niche apps beyond Excel or PowerPoint. It’s great for manual‑type tasks, yet far from handling complex end‑to‑end workflows. Some days it still shaves off half an hour, but I’ve learned its limits.

I’ve been a fan of Claude Code, but over the past few days it turned into a nightmare. The responses are sluggish, often wrong, and the whole experience feels unusable. I’m stuck watching my coding workflow grind to a halt, feeling frustrated and worried I might have to abandon a tool I trusted. This decline is seriously damaging my productivity.