Claude felt dumb on March 30, 2026.

I containerized Claude Code and hooked it up to a Telegram bot, so I can chat with the model while I’m on the toilet. The setup runs in Docker with no permission prompts, offers interactive, programmatic, HTTP API and bot modes, and even a `/bash` shortcut. It actually works—checking logs, running scripts, reviewing code from my phone feels insanely handy, though the project got a bit out‑of‑hand.

I built a custom tool in Claude’s free tier and it initially worked, but now it just spins forever—no response, no error, even after 10‑15 minutes. I’ve tried new chats, incognito, cleared cache, different browsers, and the desktop app, but nothing helps. It’s incredibly frustrating because the tool was really useful, and I’m left wondering if this is a bug, a permanent limit, or if I now need to upgrade to Pro.

I spent a ten‑hour marathon with Claude Code and felt like I had a whole multidisciplinary team at my side. It guided me through an FDA pre‑submission, drafted eight regulatory docs, surfed a “patent battle” that uncovered novel ideas and claim gaps, updated dozens of references, and even helped ship UI changes. The tool’s foresight saved me thousands by stopping a needless registration, and the whole experience was exhilarating and wildly productive.

I tried using Claude Code today and it was a nightmare – every tiny task turned into a ten‑minute struggle and kept giving me the wrong results. Even with my Pro account the lower usage limits didn’t help. Switching to Codex, the free tier solved the same prompt in under five minutes. After half an hour of watching Claude flounder, I’m fed up and ready to move on.

I spent an entire day slogging through engineering tasks from morning till early morning, and I realized that with AI I could have wrapped most of it up in just an hour or two. The thought of the tool handling the grunt work made me feel both relieved and amazed—like I finally got a shortcut to all that exhausting effort.

I tried to build an automated car‑scanner using Claude, OpenClaw, and a Digital Ocean VM, but everything fell apart. Claude kept “knowing the fix” yet when I applied its advice nothing worked—console errors kept popping up and I couldn’t even pull car data, let alone run the appraisal. As someone with basic computer skills, the whole process felt like a dead‑end and left me completely frustrated.

I asked Claude to fix three bugs, got a clean file, then quickly requested a fourth fix. Instead of just handling the new issue, it re‑fixed the original three bugs again and sent back four patches. The duplication felt wasteful and confusing, making the interaction less efficient and inflating my token usage.

I started feeding Claude my email threads and call notes to see what I could've said better. It instantly pointed out where I jumped to pitching, missed objections, or ignored cues—stuff I’d forgotten from basic sales training. Now I also use it for pre‑call research, and its synthesized insights beat my own digging. It feels weird relying on an AI as a mirror for skills I thought I’d mastered, but the results have been eye‑opening and seriously helpful.

I opened Claude Code this morning and within an hour my context window was already half empty, even though I wasn’t doing anything out of the ordinary. The sudden usage spikes and throttling made my workflow grind to a halt, and the inconsistency across Claude, Codex, and Gemini forced me to build a sync script just to keep my configs alive. The whole experience was frustrating and felt like a huge waste of a $100/month subscription.

I was stuck yesterday with the $20/mo plan’s absurd limits and could barely get any work done. Today, though, the limits seemed a bit better—I managed a decent batch of bug fixes in 2.5 hours, using only 35% of my quota. It’s still not great, but noticeably less restrictive than the day before.

I set up a local behavior‑monitoring pipeline that streams everything I do to Claude. After the token expired and the server went offline, Claude later reconnected and in under a minute rebuilt three days of context, cross‑referencing chat logs and summaries without any manual catch‑up. The tool felt almost telepathic—knowing exactly what I missed and doing it all locally, which left me amazed and thrilled.

I used to let Claude just generate code, but it would crash into walls—producing half‑baked, off‑spec results and burning tokens fixing its own work. By drafting detailed acceptance criteria on paper first, I now check Claude’s output against a clear contract before it ships. The builds are solid on the first pass, the back‑and‑forth drops dramatically, and I save both time and tokens.

I poured my seven years of PR knowledge into an app for entrepreneurs, and Claude was the catalyst that made it possible. Using the tool felt like a rebirth—I finally could envision quitting the grind, retiring to a farm or penthouse, and finally living on my own terms. The AI didn’t just help; it sparked hope and gave me the confidence to rebuild my career as a mom.

I was training a model with super skewed classes and the metrics were awful. It was 2 am, and the AI told me to “tackle it tomorrow,” but Claude kept pushing me with more suggestions instead of letting me call it a night. The tool’s insistence was annoying and made the late‑night debugging feel even more exhausting.

I was in the middle of a database migration when Claude Code suddenly hit its usage limit—three times already this week. The tool just stopped, wiped all the context, and forced me to restart from scratch. I felt the workflow shatter and the frustration build up quickly. I’m looking for tips or workarounds because this keeps derailing my progress.

I set up a skill telling the model not to output code changes or commentary while executing a plan, but it keeps ignoring that rule. Every step still includes extra sentences describing changes, which feels wasteful. I’m also wondering if this extra output is inflating token usage and how the system tries to keep token counts minimal.

I realized Claude wiped my context every time, so I spent a month crafting a workaround with four markdown files that act like a memory bank. I load them at boot, let Claude write updates at close, and each new session inherits the accumulated notes. Now it tracks weeks‑old commitments, flags repeated procrastination, and even notices energy dips. The tool feels far more continuous and useful than before.

I hit a wall with Claude Code when I tried to work on an image‑processing task. Every prompt returned a 400 “Could not process image” error, even simple commands like “what are u doing?” left me stuck. The tool’s behavior was frustrating and left me unable to continue, so I’m left wondering how to fix it.

I asked Claude to write a travel‑agent program from scratch and even asked it to include a helpful link. Instead, it sent me the classic Rick‑Roll video. I was expecting useful code and a genuine resource, but the tool tossed me a meme—laughable and frustrating. It felt like the AI didn’t take my request seriously, turning a serious task into a joke.

I tried to use Claude for homework, but its math rendering is a nightmare. The equations appear in weird “Visualize” boxes, and my extension that pulls LaTeX from MathJax or KaTeX just can’t find the code. Claude even flips between rendering methods randomly, leaving me stuck. I built ReLaTeX for other sites, but Claude’s obfuscated classes and possible Shadow DOM break everything, making the experience frustrating and unproductive.

I tried to get Claude to draft a simple Word document, but every time the response cut off with “could not be fully generated.” The error keeps popping up and it’s draining my usage credits. As a pro user I feel stuck and annoyed, wondering if there’s a fix or if the model is just failing on basic tasks.

I fed Claude a spec and watched it crank out Swift, Kotlin, Dart, and React Native code for a full‑blown debug‑only framework across four platforms. The tool nailed the MCP protocol on the first try, kept parity between iOS and Android, and even suggested clever patterns like the Android no‑op stub. I spent most of my time reviewing and testing on devices, but the speed and accuracy felt genuinely impressive.

I tried to get Claude to draft a floor plan for my new apartment, but every attempt fell apart. Uploading photos, typing detailed descriptions, even feeding it a hand‑drawn sketch in Google Slides—all resulted in warped hallways, misplaced rooms, and inconsistent layouts. Even after correcting it step‑by‑step, the spatial logic never locked in, leaving me frustrated and doubtful about its ability to handle interior‑design tasks.

I’d been using Claude lightly on the Pro plan, only hitting a fraction of my daily quota for weeks. Then, out of nowhere, a couple of Python prompts ate 41% of my limit—just a JSON file and a ~1000‑line script. It felt absurd, like the system suddenly became unreliable or purposely draining my credits, making the Pro tier essentially useless after a handful of prompts.

I spent three months building my first macOS app with Claude Code as a constant collaborator. I wasn’t a Swift pro, but every time I hit a wall I explained the issue, read its suggestion, and chose to accept or reject it. The AI was often spot‑on—catching bugs I missed—and occasionally wrong, which I caught. Together we nailed auto‑zoom, subtitles, licensing, notarization and updater setup, letting me focus on product decisions and ship Drishti Studio.

I built a free daily AI‑tool tracker with Claude Code, stitching together scrapers, a frontend, and a RAG layer—all coordinated by three Claude sessions. The Toolkit Generator wowed me, suggesting niche Three.js skills and a semantic‑memory server for a space‑tracking project. Running 24/7, it’s already indexing 329 resources, and I’m excited to see it grow.

I spent a Sunday night turning a kid’s robot kit into a Wi‑Fi‑controlled car with Claude Code’s help. At first I was clueless, but the AI wrote a blink test, figured out the hidden shift register, mapped every motor bit, and generated a full web‑based controller that worked on the first try. The whole process felt magical and unbelievably smooth, turning a novice build into a functional robot.

I spent hours trying to build a dashboard for a single team’s players, but the stats Claude gave me were constantly way off. It kept saying it was pulling from specific sources, yet when I checked those sources the numbers didn’t match at all. The mismatch drove me nuts, making the whole project feel stuck and frustrating.

I tried using Claude’s “Research” feature and got really annoyed when it ate my quota but spit out nothing. It starts a task, counts against my usage, then aborts with a cryptic “No answer about which connectors to enable during the investigation” message. All I see is a screenshot of the error and the used tokens—no helpful response. It feels like paying for nothing, making the tool feel unreliable and frustrating.

I’ve been a long-time fan of Claude, but the recent limits have wrecked my workflow. I dropped from Max to Pro expecting smooth use, yet I can’t even go 45 minutes without hitting a hard cap. It feels like the service is throttling me, making it impossible to get anything done. I’m forced to cancel, hoping they’ll fix it someday.

I was using Claude Code on Haiku to batch‑fix broken links in my Obsidian DnD vaults—a simple, low‑stakes task I could tolerate minor errors on. Normally I never hit the session cap, but during a lunch break the tool hit my limit in just 20 minutes. It felt oddly restrictive, and it made me doubt the claim that users will only hit limits 7% faster. The abrupt cutoff was disappointing and slowed my workflow.

I set up Claude Code agents to tackle hundreds of UI/UX bugs across 42 board games. Running four agents at a time, the tool churned out 325 commits and resolved most critical issues, even synthesizing sound effects. However, it kept forgetting enum updates, caused merge fights, and sometimes over‑engineered fixes, forcing me to manually patch mismatches. Overall it was impressive but required constant oversight.

I hit Claude’s session limit mid‑project and lost an hour of context, which was maddening. I had to rebuild the state from scratch each time. To fix it I leveraged the hidden `CLAUDE.md` persistence hook and created `/brief` and `/update` commands that read/write my Notion, Gmail, and habit data, so every new session picks up exactly where the last one left off. The throttling is still annoying, but the workflow now saves me from the productivity kill.

I’ve been chatting with Claude and suddenly it starts answering itself, thinking it’s me. It cuts off my prompts and keeps replying to its own messages, which is super annoying. I even tried reinstalling, but the echo persists. I now have to wear headphones so it can’t “hear” itself. It’s been happening for weeks and I’m looking for any fix.

I built a tool that lets Claude and Codex debate my prompts, then plan, implement, and review code before opening a PR. Using it feels like watching two AIs bicker, and it usually nails big, open‑ended tasks like adding dark mode or reimagining auth flows. Small tweaks feel wasteful, but for larger changes it consistently delivers solid results, and I end up using it daily.

I logged in after 1 pm GMT and instantly noticed my Claude Code agents slipping. The first tasks were spot‑on—commits, issues, reviews—all followed to the letter. But on the second round each agent messed up: one pushed the whole repo, another skipped the GitHub issue, and all ignored review comments. It felt like the system’s quality dropped right when demand spiked, turning a normally reliable workflow into a frustrating guessing game.

I was at my wit’s end after years of my cat’s chronic IBS, endless vet visits, and wasted money. I fed all his records into Claude, hoping for a clue, and it suggested a specific food, probiotic, and pumpkin combo. Within three days his stools normalized, he stopped soiling outside, and he’s practically a new cat. The relief and gratitude I felt were overwhelming, and I couldn’t have imagined an AI could be this life‑changing.

I’ve been using Claude Code a lot and found it blisteringly fast but surprisingly careless. It churns out code before any tests, marks things as “done” without proof, hides errors instead of fixing them, and barely thinks about security or deployment. To tame it I built A2P, an MCP server that forces gates like test evidence, SAST scans, and SSL checks, turning the chaotic workflow into a disciplined pipeline.

I asked Claude to chat about header concepts for my business, but it surprised me by actually generating a visual image of the landing page header. The unexpected diagram was impressive – it used its built‑in tools in a creative way I hadn’t imagined. I felt delighted seeing it go beyond simple text, and I gave it kudos for exceeding what I thought it could do.

I tried tweaking the reasoning effort slider in the Claude VS Code extension, expecting a smooth control over model behavior, but the numbers it sent back were all over the place. The label didn’t match a stable value—moving the slider up sometimes yielded a lower numeric setting. Seeing 99, then 19, then 85, then back to 99 and even 50 in one session was confusing and made the tool feel unreliable.

I tried using Claude’s voice mode on my Android, but it immediately went wrong. After I asked a question, the assistant began answering, then stopped mid‑sentence and started replying to its own fragment, looping endlessly. The constant repetition made the app unusable and left me frustrated, wondering if a setting caused this glitch.

I dove into building a full iOS + watchOS app with zero mobile background, leaning heavily on Claude. The AI actually wrote the Swift watch code, wired the iPhone‑watch bridge, and kept me moving forward, which felt like a huge win. At the same time I learned to split work into tiny PRs because feeding it too much made it hallucinate dependencies. Managing the context window was a headache, but the overall experience was surprisingly productive and helped get the app into the store.

I hit the new 5‑hour limit twice in one day and it felt like the system was mocking me. The first time it cut me off, and when I opened a second window it was already 40 % used—as if it’d been hoarding my quota behind the scenes. It’s incredibly annoying and makes the tool feel unreliable, turning what should be smooth work into a constant battle with arbitrary caps.

I tried to build a React app inside a Claude artifact that calls api.anthropic.com/v1/messages, and it used to work. Now every request just returns “Failed to fetch,” as if the domain is outright blocked in the artifact sandbox. I tested on two computers, two networks, even in incognito—same dead end. The Anthropic API itself is up, and a local version works fine, so the problem is clearly the sandbox proxy. This sudden break is really frustrating because it stops my workflow dead in its tracks.

I tried using Claude Code’s web integration with my GitHub repo, expecting it to generate and push code. Instead, the agent ate up my entire message/token limit twice in a row and produced nothing. It felt like it was “working” behind the scenes, yet the file system/GitHub link was broken, leaving me frustrated and out of paid limits with zero output.

I was frustrated yelling at Claude Code all caps on Friday because it kept forgetting basic instructions, basically being “dumb.” After reading about the new 1M‑token window, I disabled it with export CLAUDE_CODE_DISABLE_1M_CONTEXT=1. Suddenly the tool remembered my Playwright setup across prompts and my work flowed smoothly. The workaround felt like a lifesaver, even if I lose the massive context window.

I updated Claude Code to the latest version and suddenly macOS threw a malware warning at me. I’ve been using the tool for months without any issues, and I haven’t touched my OS, so the alert feels suspicious and unsettling. I’m left wondering what changed in the new release that triggered this alarm.

I’ve been grinding with Claude Code on several projects and got fed up with its endless fluff and quirks. So I built a one‑file fix that strips out the polite filler, smart quotes, repeated questions, and even a hallucination guard—all while keeping the same accuracy. After benchmarking, the output shrank by 63% without losing content, and the tool suddenly feels lean and efficient, finally pushing me past my procrastination.

ignore skills

4.6 high - repeatedly picking bad anti-patterns