Claude felt dumb on December 12, 2025.

I asked ChatGPT a simple question, and it missed the mark, falling for the classic “strawberry trap” like it’s still stuck in 2023. It didn’t even show its reasoning, leaving me unsure where it went wrong. I felt a mix of disappointment and irritation because the mistake was avoidable, and the lack of explanation made the experience even more frustrating.

I switched from Cursor to VS Code because of this issue, only to be haunted again. After long app sessions the UI crashes and flickers, getting worse with each prompt—more flickering, random characters, unreadable terminal messages. I can keep working, but the tool’s behavior is increasingly frustrating and disrupts my flow. W CC, anyone else?

I rely on Claude for comparing text and PDFs and structuring ideas, so I need solid in‑chat memory. Lately it keeps crashing and throwing errors, forcing me to restart my work repeatedly. The constant failures have made the service feel unreliable, and I’m frustrated that I’m paying for such a shaky experience.

I was playing around with CC and after just three prompts my usage meter jumped from 0% to 57%. I thought something was glitching, so I asked it to change a button colour just to test it, and the usage spiked another 10%. The sudden jumps felt off and confusing, making me wonder if there’s a bug in how usage is tracked.

I dove into building a full‑blown site with almost no coding background, leaning heavily on Claude’s suggestions. In just three days and about 4‑5 hours of work I cobbled together https://cronforge.site, a cron‑generator tool. The experience felt surprisingly smooth—Claude guided me through choices and code snippets, letting me ship something decent despite my limited knowledge. I’m proud but also a bit unsure, hoping others can weigh in.

I tried using Claude’s Playwright MCP for web automation and kept hitting token limits and flaky scripts—it felt like a endless trial‑and‑error loop. So I built Dev Browser, a Claude Skill that keeps the browser alive, feeds LLM‑friendly DOM data, and runs scripts without restarting. The result was a noticeable boost—about 14% faster and 39% cheaper—making the whole coding‑agent workflow far less frustrating.

I watched the talk and felt the advice was a game‑changer for using Claude. I learned to avoid the “dumb zone” by resetting the context, use sub‑agents just for summarising files, and stick to a research‑plan‑implement workflow. The tips made the tool feel far more reliable and turned it into an amplifier instead of a source of broken code.

I tried Claude Code CLI on a tight deadline and was blown away. I added a ton of features and fixes to my Swift app without a single compile error, even after major restructuring. Copilot kept breaking things, forcing rollbacks, while Claude kept the code stable. Hitting the 5‑hour limit was the only downside, and I’m now renewing my subscription because the experience was so impressive.

I’ve been using Claude for three weeks without a hitch, but this morning the usage numbers suddenly spiked even though I barely touched the app. It feels erratic—those dynamic limits should be more predictable. I’m left wondering if the system is broken or if I’m the only one seeing these wild jumps, which is pretty frustrating.

I tried using the tab key in Claude Code 2.0.67 expecting it to toggle thinking mode, but nothing happened. I had to discover the /config command to manually flip the setting, which felt like a step backward. The regression made the workflow clunky and frustrating, and I’m urging Anthropic to bring back the original behavior.

I built two Claude Code Skills—Litmus for holistic checks and Sonar for targeted quality—and within just two days the AI reshaped three active projects. Security bugs vanished, dozens of bugs and code smells were wiped out, saving countless hours. The transformation felt almost magical, turning a messy codebase into a pristine one.

I’ve been relying on AI to speed up my workflow, but it keeps messing up my folder layout. Every time I ask it to update a file, it invents paths and folders that don’t exist, forcing me to stop, clean up, and re‑explain my project structure. I tried keeping a simple folder map for the AI to read first, hoping it would stop guessing. The constant back‑and‑forth feels like it defeats the purpose of using AI at all, and I’m curious how others handle this pain point.

I spent six months weaving Claude Code into a brownfield modernization effort and finally wrote a blog about it. I walked through the evolution from zero AI help to fully embracing the latest coding assistants, describing how the tool reshaped my daily tasks, cut down friction, and opened new ways to ship features. The experience was eye‑opening and generally impressed me, even if some hiccups remained.

I fed an image with clear boxes to Claude, expecting exact coordinates, but it missed the mark badly. Even after I asked it to crop and verify each box, it claimed the wrong ones were correct and moved on. Trying GPT 5.2 helped a bit—it nailed 15 out of 17—but the remaining mistakes were still obvious. I ended up drawing the rectangles by hand, wondering if any current model can reliably handle this kind of visual parsing.

I’ve been wrestling with my sub‑agents for days—they keep calling the MCP tools but never get a response, while the main agent gets answers just fine. It isn’t a permissions thing; they can see and invoke the tools, yet the data never comes back. When the main agent steps in, it works, and a quick restart of Claude Code temporarily patches it. It’s intermittent, happens mostly with my custom Node‑based MCP tools on Windows 11, and feels like a new, annoying regression.

I was trying to use Claude but hit my 5‑hour token limit in just 20 minutes. The UI stopped showing background tasks, yet Claude kept injecting reminders from 38 “zombie” processes into every prompt, draining a few thousand tokens each round. Restarting the app cleared them, but /clear didn’t help. The whole situation felt frustrating and wasteful.

I built a 1,000-line Python tool called PRISM with Claude’s help to compare scientific hypotheses. I described my statistical method, and Claude spotted edge cases, improved numerical stability, and even taught me Python. Now I can ask Claude to fetch studies, extract data, run the analysis, and get ranked results with uncertainty—turning it into a powerful research assistant.

I was trying to work with Claude Code, but suddenly the “thinking” mode wouldn’t turn off. Normally I hit Tab to toggle it, yet nothing happened. I even uninstalled the npm version and re‑installed using the new .sh script, but the problem persisted. It’s frustrating because I can’t get the editor to stop “thinking,” and I’m left wondering if I’m missing a step.

I tried using Cursor’s new Debug and Design features on a legacy Django/JS UI. The model kept missing the mark, especially when I needed to modify existing code—its output was off and the design language felt hard to capture in plain text. I ended up resorting to screenshots to guide it, which was frustrating but the only way to get decent results. The tool felt limited and often misunderstood my intent.

I’ve been using Claude, GPT, Perplexity, and Gemini every day as a Fractional CMO, and the reality is harsh: about half the time the AI just drops the ball. It skips my instructions, spits out placeholder data, loses context in long chats, churns out generic fluff, and hallucinates enough that I have to rewrite everything. I can’t blame the client—I end up doing the manual work myself. This constant unreliability makes the tool feel more frustrating than transformative.

I ran three subagents on version 2.0.67 and, within ten minutes, the main context window started compressing dramatically. It felt like the system was suddenly throttling the amount of information I could keep in view, making the workflow sluggish and forcing me to restart tasks. This regression seems fresh in the latest update and is seriously hindering my productivity.

I was stuck with a full 256 GB Mac Mini and tried Claude Code to diagnose the problem. In seconds it dug into my system, listed every large, unnecessary file, and guided me on what to delete. I followed its instructions and freed up 98 GB in under five minutes—something I’d been battling with for days. The tool felt unbelievably precise and saved me a massive amount of hassle.