Claude felt dumb on April 11, 2026.

I tried Claude back in January to build a headless orchestration tool for a coding workshop. At first it ran four instances, handled memory and chat MCPs, and I was convinced it deserved a $100/month plan. Now I can’t even get a single direct‑interaction instance running, the response quality has plummeted to a “drunk junior” level, and I was never warned about token limits, model changes, or peak‑hour restrictions. The silent downgrade feels like a scam, so I’m looking at local models like Hermes + Ollama instead.

I built a notification add‑on for Claude Code and after 374 downloads I finally stress‑tested it. I discovered every alert just said “Task complete” because the parser couldn’t read Claude’s nested tool calls, so the summaries were missing—a frustrating oversight. After fixing the JSON handling and silent HTTP errors, I released v1.1.0 with worktree‑based IP isolation, and now the tool works as promised.

I was fed up with people dismissing my observations, so I dug into Claude’s effort tags using my company account. The output clearly showed distinct <reasoning_effort> values for low, medium, and high settings, and “max” removed the tag entirely, matching the docs. The whole ordeal left me irritated by the gaslighting and eager to prove the tags really exist.

I noticed that when I run Disbatch/Cowork with the Claude desktop tab open, my CPU spikes dramatically, even though only one Claude code is active and no heavy tasks are running. Strangely, minimizing the tab cuts the usage roughly in half. The whole situation feels odd and frustrating, making me wonder if it’s a bug or something wrong with my computer.

I was grinding on late‑night homelab maintenance, rebooted a node, and typed `claude -c` out of habit. To my surprise the conversation kept going, synced via Syncthing across the cluster. I ended up with a repo, a setup script, and later another node picked up the same project and finished it. The tool’s seamless session continuity felt oddly magical and saved me a ton of re‑typing.

I grabbed Google’s AI Edge Gallery, pulled the Gemma‑4 weights, and tested it offline on my flagship phone. The 2B model felt snappy, the 4B was usable but quickly heated up and drained the battery. Multimodal image ID worked well, audio was passable. Compared to a MacBook Air it lagged, but the privacy‑first, zero‑API setup impressed me, even if older phones will choke.

I was building a simple betting‑app HTML file and everything was fine until Claude suddenly started saying the conversation was too long and wouldn’t continue. I cleared the chat, disconnected GitHub, logged out, even made a fresh account, but as soon as I re‑uploaded the same HTML the error reappeared. It feels like Claude has hit some hidden limit on how many times it will work with the same file, and I’m stuck looking for a workaround.

I built my first project, Cortex, mostly with Claude Code’s help—it actually wrote the bulk of the codebase and even became one of the agents tracking itself. I set up a task chain where agents must get approval before closing tasks, using Claude as builder, Codex as reviewer, and Hermes/GPT for research. The experience felt empowering, and I’m excited to share the open‑source repo for any feedback.

I was on my Android using Claude and discovered the new “Dispatch” feature, which let me control the Claude UI on my Mac. After a memory‑leak crash and a reboot, I gave the permission and asked Claude—remotely—from my phone to scan the kernel logs. It pinpointed the culprit (an MCP Server in VSCode), suggested a workaround, and even helped me draft a clear bug report. The experience felt surprisingly smooth and useful.

I watched the system eat an entire day’s Pro‑plan quota in six minutes, and it was maddening. I kept shouting “STOP” as it bulldozed my project, spitting out a needless binary wrapper, pulling irrelevant research, and then rewriting my simple 25‑line plan into a massive Python client. The tool’s behavior was chaotic and wasteful, leaving me frustrated and questioning the product.

I’ve been running my whole daily workflow through Claude Code for a month, and it’s become my personal OS. I set up two projects—one for personal life, one for two jobs—linked via a shared Obsidian vault and semantic search. Mornings start with a briefing that merges emails and tasks, I control my phone over Wi‑Fi, and a Power Automate bridge handles corporate email. The system auto‑journals my chats, keeps a three‑layer memory that survives resets, and learns from every failure by turning it into a permanent rule. It’s not perfect, but it’s transformed how I manage everything.

I upgraded to the $200 plan expecting a smooth experience, but Claude Code just froze on a simple bash task for over 40 minutes. While the browser, desktop, and mobile apps worked fine, the coding assistant wouldn’t respond at all. It felt like a broken promise—like I’d paid for a tool that was completely useless in that moment.

I spent a few days tinkering with Claude to create an equalizer for Safari because watching videos at 2× speed mashed vocals together. Claude guided me through the code, and the result lets me boost volume, isolate channels, and apply presets without extra drivers. After battling four Apple rejections, the app finally went live, and it’s been a game‑changer for my productivity and listening experience.

I’ve noticed the model feels sluggish now – the same tasks need several prompts, retries, and extra clarifications. What used to be one clean answer now burns a lot more tokens, forcing me into back‑and‑forth iterations. It’s not broken, but the drop in efficiency is frustrating and feels like a hidden limit cut.

I’m blown away—after a single day’s worth of work I hit the usage limits in minutes. I’ve never run out so fast, even with the max 20x and 5‑hour caps. The tool cranked out results so efficiently it felt like Anthropic was pushing my limits on purpose. I’m literally speechless.

I use Claude daily and keep hitting a maddening roadblock: it never knows the current time. I get “go rest” at 2 pm, “call the doctor” at night, and useless scheduling hints because the model lacks temporal context. I’ve even built a simple script to inject timestamps, proving it works, yet Anthropic hasn’t implemented it, while competitors already have clocks. The missing feature feels like an avoidable oversight, making the experience consistently frustrating.

I keep battling Claude’s Chrome extension—it feels like a never‑ending tug‑of‑war. Every time I try to get it to help, the tool misbehaves or drops my prompts, forcing me to abandon it and fall back on manual console logs. The friction is tiring, and I’m still unsure if it’s worth the hassle.

I kept trying to get Claude Cowork to click simple buttons in Chrome, but it constantly missed or failed at random. Every time I needed it to press a “Submit” or “Next” button, I had to watch it try again and again—often five attempts just to get one click. It felt exasperating and wasted a huge chunk of my time, making the whole workflow feel stuck.