Claude felt smart on April 18, 2026.

Полный кретин. Не способен ни системному промту следовать не слышит команд в чате. Просто пиздец. Тариф MAX X5 100$

I keep getting hit with “API Error: Server is temporarily limiting requests (not your usage limit) – Rate limited” every few minutes, which makes it impossible to keep working. Since the 4.7 release, Claude’s code generation has also felt off, producing buggy snippets that need extra fixing. The constant throttling and the dip in quality are really frustrating and slowing me down.

I was shocked when Claude Code started erasing my organization’s .claude/settings.json every time there was any mistake in our custom marketplace folder. I’d catch the error messages only once, then everything vanished, forcing me to rebuild the settings manually. The whole experience felt unsafe and incredibly frustrating, especially since it wiped crucial config data without warning.

I kept getting stuck in an endless login redirect loop with Claude Design despite being a long-time Max x5 subscriber and having my AdBlocker turned off. Even logging out and back in didn’t fix it. The whole experience was maddening because I couldn’t even access the service to get any work done, and I felt completely blocked by this broken behavior.

I kept Claude Code running and watched it gobble RAM like it hadn’t eaten in ages. My 16 GB machine started freezing and eventually crashed because the app filled up memory. It made the whole PC unusable, and I’m left wondering if the Claude‑mem module is the culprit. The experience was maddening and completely broke my workflow.

I switched from Copilot Business to Claude Max and hit a wall trying to recreate my old seamless agent plan review workflow. The tool forces UNC path approvals constantly, even in “yolo” mode, and planning now needs way more context. I can’t see changes without committing each time, and there’s no way to set global instruction sets for languages like /cpp or /python. All these hurdles make the experience pretty frustrating.

I set up Claude Code to run a /loop that checks my Linear board every ten minutes, pulls tasks labeled “Claude,” and lets the AI draft emails, research, and generate content automatically. I can tweak its output by commenting on the issue, and it iterates right there. The whole workflow feels like async pair‑programming right inside my issue tracker—transparent, auditable, and surprisingly smooth, even if the session dies when the cron stops and token costs can climb.

I spent 15 minutes uploading my design system to Claude Design, only to have my entire week’s usage nuked instantly. The abrupt cutoff was infuriating, effectively halting any progress I had planned. I’m pleading with Anthropic to add a feedback‑reward system during this preview so that giving input actually earns us back the usage we lose.

I spent the day testing Claude Design on my usual dashboarding tasks and was genuinely impressed. It handled visual analytics smoothly and felt like a natural upgrade from the Figma plugins I use. The seamless integration with Claude Code made the workflow feel faster and more cohesive, and I can already see it replacing my current tools. The experience left me excited about how analytics pros could adopt this in their work.

I’ve been testing 4.7 and it feels like a side‑grade rather than a true upgrade. It’s sharper on a few tasks but actually worse on others, and the new usage limits and the delayed Mythos release just add to the disappointment. It makes me wonder if we’re staring at a ceiling in LLM progress.

I was constantly re‑loading context for Claude each morning, wasting 20 minutes just to get it up to speed. After I stopped treating Claude as the product and built a memory stack around it—code‑repo awareness, semantic Neo4j graph, session logs, Obsidian notes, custom skills—the model started recalling past decisions and catching patterns I’d forgotten. Now it feels like a coworker with real memory, not just a smarter answer bot.

I used Claude Design on Pro to overhaul a mobile flow in a Next.js app, and it delivered a solid first draft in minutes. I scoped the exact files, got a concrete audit plus TSX changes, exported a ZIP, and merged the code locally. The tool was fast and precise, though I still did local QA. Overall the experience felt efficient and reliable, making the redesign feel much less painful.

I recorded my 6‑year‑old’s invented “Bloop” language and fed it to Claude, asking if it was real. The model returned a 47‑word vocab, three grammar rules, and even noted a negation pattern similar to real languages. When I tested it on my son, he froze, switched to English, and looked betrayed. The whole experience was startlingly accurate and left me both amazed and a bit unsettled.

I finally have the mentor I always wanted—Claude’s teaching mode walks me through Linux command line tasks in Zsh, pointing me at the actual source on GitHub and even the main.rs file. I felt the tool’s guidance was spot‑on, turning a daunting learning curve into an interactive, confidence‑boosting experience.

I tried using Claude’s latest CLI, but I keep hitting version 2.1.88 and can’t get the source code for updates. Every time I run it, the responses are filled with nonsense, so I’ve had to cobble together a supervisor system just to filter out the BS. The whole experience felt wasteful and really frustrating.

I spent a few evenings prompting Claude and ended up with a fully‑functioning app hooked to Supabase, all without any coding background. The experience felt surreal—what would have cost tens of thousands from a contractor was done by a chatbot. I’m thrilled with the results, even though rate limits bite a bit, and I’m now wondering if there’s any hidden downside.

I was surprised when Claude suddenly dropped a swear word. I’d even asked it to flag terrible or stupid ideas, and it usually helped me brainstorm or study. Hearing it curse was jarring and a bit disappointing, making the experience feel less polished than I expected.