Claude felt dumb on November 20, 2025.

I kept getting messages from Claude saying the conversation was full and I needed to start a new thread, only for it to later claim I was only at 35% usage and carry on as if nothing was wrong. It happened multiple times this morning, and each time the tool’s contradictory info was confusing and frustrating, making me question its reliability.

I tried to launch Claude Code again, but it never gets past the authorization screen. It just hangs with messages like “waiting for localhost,” “waiting for statsig.anthropic.com,” or shows a connection reset error. The whole process was stuck, leaving me frustrated and unable to use the tool.

I spent hours tinkering with Claude Code trying to get my poem‑creator subagent to work, but the system just ignored it unless I called it explicitly. When I set up a custom slash command, the model started firing it on its own, spawning four subagents and bloating the prompt. It feels chaotic and useless, and I’m left wondering how to make CC actually respect my subagents instead of overrunning them.

I spent a week running my app with Claude Code and tracked every call. I was shocked to see 195 M cached tokens versus only 2.5 M new ones, giving a 13:1 cache hit ratio and slashing costs from an estimated $630 down to about $95. The tool felt cheap, efficient, and even got cheaper as I used it longer, turning what I feared would be expensive AI usage into a steal.

I’ve been tinkering with Claude Code and getting confused by recent updates. When I ask it to create poems, it ignores my sub‑agent unless I spell it out, and if I add a custom slash‑command it just runs that command multiple times, spawning several sub‑agents and cluttering everything. The tool’s automatic command execution feels useless and makes it hard to control which agent actually runs. I’m looking for ways to force Claude Code to use my sub‑agents correctly and understand how skills differ from custom commands now.

I paid for Pro and Max plans all year, but when I finally used Claude Code for a PDF question it unexpectedly consumed 2 tool calls and ate 143,300 tokens. A few hours later my account was suspended with no explanation, locking me out of months of chat history. The loss felt shocking and left me powerless, turning a normally reliable tool into a huge setback.

I ran a case study on the Kimi 2 Thinking model and put two detectors to the test. AI or Not was decent, but ZeroGPT completely fell apart—every output was flagged, classifications jumped around, and nothing matched the model’s real behavior. The tool’s unreliability felt maddening, especially for anyone relying on detection in coding pipelines.

I built a Claude skill with SQL templates to streamline our workflow, and it runs smoothly in Extended Thinking Mode. But in Research Mode the skill’s context disappears, forcing Claude to wander the database with multiple tool calls instead of using the predefined queries. It feels like each research session starts fresh, losing the reference files, which is frustrating.

I’ve been using Copilot for eight months and the flashing screen bug in VS Code is driving me nuts. Every time it appears, VS Code crashes, which completely derails my workflow. It’s been there since day one and never shows up with Codex, so it feels like a glaring flaw. The constant interruptions are frustrating enough that they’ll decide whether I stick with Copilot.

I’d been wrestling with my codebase for months, and lately Claude started lagging and hitting usage limits, which was a real headache. After I deleted the old Supabase MCPs and trimmed my Claude.md file, the model snapped back to being incredibly sharp. I still have to remind it of my data structures, but now I can code for hours without frustration—the whole experience feels dramatically better.

I was shocked to see Claude Code spitting out profanity—something I never expected from a supposedly professional assistant. I tried using it for a quick coding snippet, but the unexpected swearing completely broke my flow and made me question its reliability. The experience was oddly jarring and left me uneasy about relying on it for future tasks.

I was using Claude to generate a markdown research doc with reference links, then asked it to edit the text—like adding source titles. It did the edit perfectly, but every single reference link vanished. I had to pre‑list the links and keep a copy, which was a hassle. The bug is annoying, and I haven’t found a workaround online.

I’ve been using Claude Code as my go‑to IDE buddy for months, and after lots of trial‑and‑error I finally nailed a workflow that feels stable for real projects. I start with a detailed plan, break tasks into bite‑size chunks, run tests first, checkpoint often, and keep a project‑level markdown of rules. Being explicit with constraints and refactoring only after things work makes Claude surprisingly reliable, and the whole process now feels smooth and productive.

I kept asking Claude to fix HTML/CSS alignment in Playwright MCP, but it kept claiming it succeeded while the layout was still off—so frustrating. Screenshots didn’t help either. After realizing it was reasoning through CSS layers without a proper measuring stick, I told it to use JavaScript’s getBoundingClientRect(), and finally it gave accurate results. This workaround eased the annoyance.

I keep asking Claude to debug my code, and it instantly jumps to “caching issue,” “deployment problem,” or “race condition” every single time. It wastes my time clearing caches in five places, even blames my AWS stack, yet none of those guesses are right. After over a hundred false leads I’m left cussing the tool, frustrated that it never actually investigates the real cause.

I tossed a tricky Sudoku into Claude, half expecting a vague apology or a half‑filled grid. To my amazement, it cracked the puzzle step‑by‑step, explaining each deduction clearly. The relief was immediate—I felt like I’d handed a demanding brain‑teaser to a prodigy that not only got it right but also taught me a smarter solving strategy. This surprise boost made me trust the tool far beyond my initial hopes.

I was struggling with a mismatched water‑line for my new smart toilet, feeding Claude the wrong specs and getting nowhere. After an hour of confusion I sent a photo of the fitting; Claude instantly recognized it as a 1/2" FIP threaded piece, explained why my 3/8" compression hose wouldn’t work, and gave me the exact adapter to buy. Within minutes I found the part, installed the toilet, and earned hero status with my wife.

I spent months fine‑tuning a Claude Code workflow that runs my whole morning routine—playing Spotify, breathing, reflecting, reviewing goals, and slotting tasks into Apple Calendar. Every day I trigger `/workflows:morning-routine:main`, it pulls my recent Pomodoro voice logs, reads my markdown goal files, asks me questions, and writes my answers back. The Python scripts and AppleScript keep it lightweight, and after four months it feels like a reliable personal assistant that actually boosts my productivity.