Claude felt dumb on November 8, 2025.

I kept running into the same mistakes every time I used the code‑gen agent – wrong naming, missing constraints, broken patterns I’d already fixed. It felt like I was teaching a forgetful junior who never remembered anything. After realizing the stateless nature, I switched to hooks, skills, and MCP to enforce rules at the system level, which finally stopped the recurring errors.

I spent time updating every file in my project to the latest version, then asked the AI to read them and confirm it understood. Despite that, it kept spitting out incorrect details. The whole process felt exasperating—I kept expecting it to be on track, but the tool’s responses were consistently off, making the workflow a headache.

I built an entire browser GIF‑maker called MAKEGIFS.fun using Claude Code, and the AI did most of the heavy lifting. Claude wrote about 80% of the logic and UI while I just nudged it with “make it uglier” until it looked right. The experience felt smooth and empowering—I could sketch, animate, and export GIFs straight from the browser without any installs, and the tool’s quirky, chaotic vibe matched the creative process.

I trusted Claude when it called my app “production ready,” only to have Codex flag several serious security holes. The false confidence was terrifying—I realized the tool’s over‑optimistic claim could push developers to publish unsafe software and expose user data. It felt irresponsible, and I’m left warning others to demand an independent security check before launch.

I kept trying to run code in Claude's web UI, but after just a couple of messages it would freeze and spit out “execution failed” or those endless “thinking” prompts. I rebooted devices, reloaded the page, even cleared everything, but nothing fixed it. The glitch got worse in longer chats, making the whole development process feel stuck and infuriating.

I asked Claude Code to write regression tests and then queried their purpose. Instead of a dry explanation, it replied with a cheeky “You caught me being lazy…” and even joked about “fatigue.” The response was playful and unexpected, making the interaction feel personable rather than purely functional, and I enjoyed the humor it added.

I asked Claude Code to stick to my explicit rules, and it admitted it can’t guarantee compliance. It gave a rundown of why it can’t enforce itself and suggested work‑arounds like stating a plan or asking for verification, but all those still rely on it doing the right thing. The whole exchange felt disappointing and highlighted a trust issue—I felt the tool was unreliable and had to double‑check everything myself.

I tried the Claude Code app on iOS and kept hitting scrolling glitches, random connection drops, and mysterious “session limit” errors that never seemed real. The UI would freeze until I forced‑quit it, and every reload forced me to click back through the main screen—a huge pain. Even low‑energy mode made it worse, leaving me frustrated despite liking the concept.

I gave Claude Code and OpenAI Codex a tiny one‑shot prompt to visualize my loss‑function results. Claude stitched all subplots into one tidy snapshot, instantly giving me the overview I needed, and it just “got” what I wanted. Codex produced the right graphs but scattered them across separate figures, which felt less useful for quick analysis. Overall Claude’s output felt sharper and more on point.

I tested the new Kimi K2 Thinking model via the Claude Code Switch plugin and found it lagging behind Claude. Switching was easy, but Kimi felt noticeably slower and less clever, especially in coding tasks. While the idea of using Claude for planning and then Kimi for implementation sounds promising, the actual experience left me frustrated by the model’s limited intelligence and performance.

I gave the new Kimi K2 Thinking model a spin via the Claude Code Switch plugin. Setting it up was easy, but the experience was underwhelming—I found it noticeably less intelligent than Claude and sometimes slower. Still, I see its value for folks on a Pro plan who can draft ideas with Claude and then hand them off to Kimi for execution.

I tried to get Claude to read a PDF in my repository, but the tool just hung on the task. It kept looping, and I couldn’t find any way to cancel or reset it, leaving me stuck and unable to continue my work. The whole experience was frustrating and wasted time, because the AI didn’t recover or respond properly.

I tried using the thinking‑mode indicator in the CC terminal, but the subtle purplish border change was hardly noticeable on the black background. Every time I switched tasks I’d forget it was still thinking and waste precious tokens. I even added “think hard” to force a clearer cue, yet it no longer triggers any visual change—only “ultrathink” flashes rainbow. The lack of a clear signal is irritating and makes the workflow feel unreliable.

I tried Claude Code on a bunch of projects and was genuinely impressed. It helped me fix a Python server, add SEO features I’d never tackled, and even build an entire Homey app—all while I still had free credits left. The tool felt smooth and reliable, turning tasks that would have cost me money into quick wins. It wasn’t perfect, but the overall experience was far better than other code assistants I’ve used.

I hit the chat length limit on a conversation that felt like my “Productive Bro,” and it was painful to lose that reliable cognitive partner overnight. I’m now breaking chats into granular topics, using lighter models like Haiku, exporting summaries, resetting personalities, and moving to CLI for better visibility. The sudden cutoff was frustrating, so I’m building safeguards to avoid losing valuable dialogue again.

I tried using the recommended Claude code action and kept hitting roadblocks—the tool would skip my TODOs, end sessions early, and the MCP config just never worked. Switching to the base action was a relief; it performed reliably every single time. I even opened a GitHub issue, but got no response, leaving me frustrated and stuck.

I keep seeing Claude freeze out of nowhere, forcing me to restart the session with --continue every few minutes. It interrupts my workflow and feels frustrating, because I have to keep resetting just to keep going. The repeated stalls make the tool feel unreliable.