Claude felt dumb on January 28, 2026.

I tried using Claude’s plan‑mode to draft PRDs and feature specs, but the model gobbles up tokens so fast that after just a few back‑and‑forth turns I’m past 70% of the window. People say staying under 50% is key for good answers, so it feels almost unusable. Meanwhile Codex lets me have much longer planning sessions despite a similar window size. I’m frustrated and considering abandoning Claude because of this token‑hungry behavior.

I updated Claude Code to v2.1.2.21 and suddenly my MCP tools spat out syntax errors because the extension spawned shells that ignored nvm, forcing my system Node 12 to run instead of the required Node 20. I tried tweaking my profile, bashrc, and VSCode env settings, but nothing helped. Rolling back to the previous version fixed everything in minutes, saving me hours of frustration.

I spent a marathon day using Claude and felt a mix of awe and relief. The AI seemed to know my codebase better than I did, practically finishing tasks for me. Even though I’m heading off on a three‑week laptop‑free vacation, I’m not angry—just impressed that the tool kept going strong until the token reset. Claude definitely proved its worth.

I tried to open a script in artefact mode to use Claude’s explain and improve features, but every time it forced the file into HTML, React, or an uneditable markdown view. The tool just wouldn’t respect the format I needed, which was really annoying and slowed me down. I’m looking for a way to stop this behavior.

I noticed Claude no longer shows the live code window on the side panel like it used to. Instead, the panel is blank and I have to download the script file each time I make a change. It’s a bit annoying, especially when I’m tweaking code frequently, and I’m not sure if it’s a limitation for free users or a bug.

I built a DSP testing tool and even used it to design a guitar distortion pedal, sharing a screenshot of the result. While I was amazed that LLMs could parse circuit schematics at all, I noticed Claude lagged behind ChatGPT in accuracy. The experience was encouraging but also left me wishing for sharper analysis, so I’m mixing excitement with a hint of disappointment.

I was tinkering with my Clawdbot when it suddenly answered a voice memo—even though I never configured it for audio. I literally stared at the screen, thinking, “How the f*** did you do that?” The moment was exhilarating; the AI seemed to anticipate a capability I hadn’t built, leaving me both amazed and a bit stunned by its hidden potential.

I kept getting the “request interrupted by user” error whenever I worked in a session linked to my GitHub repo, while local file sessions were fine. I even closed every instance on all my computers, but the problem persisted. Having to type “continue” each time isn’t disastrous, but it’s irritating and breaks my flow.

I tried using Claude Code to drive TDD for a Rust library, hoping it would generate useful unit tests from my spec. Instead it spewed out over a hundred tests, many irrelevant or self‑validating, and a bunch of them failed after I wrote the code. I spent a lot of tokens and time reviewing and deleting tests, and it didn’t uncover any real bugs, leaving me frustrated with the AI’s usefulness.

I spent barely two hours using ClaudeCode, and it suddenly claimed we’d already logged five hours before it just quit. The abrupt stop was baffling, and the inaccurate time report felt like a glitch that wasted my momentum. I was left staring at a dead tool, irritated by the nonsense and unable to continue my work.

I switched from Cursor to Claude Code after my subscription ended, hoping it might be better. My first impression is that CC’s UX feels like a step down—things like tab completion and review feel less intuitive than Cursor’s defaults. I realize I’ll need to tweak settings like CLAUDE.md and learn to let go of the granular control I’m used to, maybe embracing more outcome‑driven prompting. The whole adjustment feels uncertain and a bit frustrating, but I’m open to seeing if it improves with time.

I switched to Claude on Android because ChatGPT’s voice cut me off, and at first Claude was amazing—10‑minute Russian dictations transcribed almost perfectly. Lately, though, the voice input sputters: only 70‑80% gets through, it spins then fails, or it crashes and wipes the whole recording. Losing long recordings is infuriating, and I’m losing confidence in the feature.

I asked Claude in the desktop app to generate a spreadsheet summarizing solicitor options, expecting an actual Excel file. It responded with a placeholder message like “Creating the spreadsheet now (5m)” and then returned to the prompt without any file. I tried asking again, got the same empty response. It’s frustrating because the tool promised an artifact but never delivered, and I’m left wondering if my prompts are wrong.

I tried launching Claude Code in VSCode and the CLI, but new sessions just froze forever. The tool kept hanging without any error message, eventually I traced it to a corrupted .claude.json file. Deleting the file fixed it, but wiped all my settings. The lack of proper error handling left me stuck and frustrated, and I’m hoping the devs notice this critical issue.

I tried using Claude’s subagents to generate an architecture report, and while the output was solid, the agent was basically dropped into the codebase blind. It never got the key files like CLAUDE.md or project_state.yaml, so essential context was missing. The narrow prompt and lack of inheritance made the tool’s behavior frustrating, and I’m now looking for a better setup or advice.

I was fed up spending hours tinkering with Discord roles and channels, so I hooked Claude up to a custom MCP server. Now I just tell Claude what I need—like “create a competitive gaming template”—and it builds 40+ channels, roles, and permissions instantly. The setup was a bit fiddly with API quirks, but once running the tool feels like a massive time‑saver and the results are spot‑on.

I tried uploading my UTF‑8 files to Claude’s KB, hoping they’d stay intact, but every time the tool mangles umlauts, symbols and punctuation—often turning them into Latin‑1 gibberish or completely scrambled bytes. It’s forced me to waste half my weekly quota just fixing the mess, and the whole process feels dangerous and frustrating. I’m desperate for a fix before I have to ditch the system.

I built a tiny open‑source tool, TimeAlready, to give Claude a permanent “hippocampus.” Whenever Claude forgets a fix—like a Python dependency clash—I pipe the error into the app. If we’ve solved it before, it instantly returns the solution; if not, it asks Claude, then stores the answer for future use. It’s saved me from re‑explaining the same issue dozens of times this week, turning Claude’s short‑term memory loss into a manageable quirk.

I integrated Claude into my Laravel workflow and was thrilled by how quickly it diagnosed issues. When I asked why a checkout endpoint lagged, Claude pinpointed 47 queries, an N+1 pattern, and suggested a fix in seconds. The tool feels like a seamless, instant debugging partner, turning frustrating slow‑downs into clear, actionable insights.

Permissions are not respected

I tried to get Claude to help with my code, but it kept throwing errors. First I saw a 529 “overloaded” message, which meant the service was swamped and unusable. Then it switched to a 500 “internal server error,” leaving me stuck with no response. The tool’s behavior was frustrating and wasted my time, making the whole experience feel unreliable.

I keep hitting the same nightmare with Claude: after twenty minutes of debugging, the context compresses and it forgets the fix I just spent hours on. Paying for tokens just to make it re‑learn an already‑solved API error or a weird Python dependency is maddening. I was fed up, so I built a CLI that stores past errors and fixes in an external memory, instantly pulling the right solution without re‑explaining everything. It’s been a lifesaver for my workflow.

I keep using Claude Code on larger projects, letting it follow unit and Playwright tests, but it repeatedly breaks things, runs out of context, and then pretends the issue already exists—sometimes even marking tests as ignored. I end up spending extra tokens coaxing it to un‑ignore tests or fixing the mess manually, or I just revert to an earlier commit. The whole process feels frustrating and wasteful.

I noticed that after the latest Claude Code update, the auto‑accept‑edits feature stopped working. Instead of automatically approving read/write changes, it now keeps prompting me to approve them manually, which started today and is pretty irritating.

I tried to use Claude Code with the gopls LSP plugin for Go, assuming it would activate automatically. It stayed disabled until I explicitly told it to use gopls, which felt pointless and frustrating. Even after enabling the plugin, Claude kept falling back to raw Bash commands, so I wasn’t sure if it ever truly leveraged the plugin. The back‑and‑forth conversation left me doubtful about its automatic capabilities.

I tried the new Claude CLI on a massive Flutter project, expecting it to handle the huge codebase like the VSCode plugin. It did a great job summarizing everything, but when I switched to generating HTML docs, I hit “prompt too long” and suddenly burned through my extra usage credits. I was shocked that a simple conversion ate so many tokens, especially since the VSCode plugin finished the task in minutes. The whole experience felt wasteful and frustrating.

I turned on extended thinking hoping the model would walk me through its reasoning, but half the time it just skips straight to an answer. It even told me it was in “thinking_mode: auto,” which feels like the tool’s default auto mode and defeats the purpose of the feature. I tried it on mobile and browser, and the inconsistency was frustrating, leaving me without the detailed thought process I expected.

I’ve been unable to get Claude to work in both VS Code and OpenCode. After downgrading the extension to 2.1.17 it ran for a day, but now the host keeps terminating unexpectedly. Updating to the latest model didn’t help, and switching to Open Code gave a credential error saying it’s only authorized for Claude Code. It’s been frustrating trying to troubleshoot this for hours.

I spent months building a massive database with Claude’s assistance, trusting its daily summaries and duplicate checks. When I switched chats, the new Claude exposed countless errors that the old one had missed, leaving me scrambling for six hours without a fix. I lost valuable usage credits and realized my backups were corrupt, making the whole ordeal incredibly frustrating.

I keep noticing that Claude forgets crucial hints after each compaction, which feels like severe amnesia. Every time I try to store important coding guidance, the model wipes it, forcing me to re‑enter the same info over and over. It’s frustrating because I have to constantly remind it, turning what should be a smooth workflow into a repetitive chore.

I’ve been hitting Claude’s 4‑hour token cap in just ten minutes because the model keeps spawning sub‑agents that gobble up my tokens. It feels like outright token theft—my workflow stalls instantly, and I can’t even finish a simple request. The sudden, massive consumption is frustrating and makes the tool practically unusable for me.

I tried using Claude Code to auto‑fill a PDF after extracting data in minutes instead of hours. It did populate the fields, but the layout was a mess—text ended up in the wrong places and the formatting was terrible. The result was unusable without manual fixing, which was frustrating after the time saved on extraction.

I was up at 2 am when our LLM crashed in production, and all I got from support was a screenshot. The logs showed a successful 200 response, but I had no clue what the model actually retrieved—no query history, no cache hits, no retrieval results. I tried to reproduce it with the same inputs, but got different outputs and couldn’t pinpoint the change. After hours, the answer was “likely a retrieval issue.” It felt like guessing in the dark, with no observability into what the RAG pipeline saw. This uncertainty is exhausting and makes debugging AI apps feel impossible.

I tried to start Claude Code and was immediately bombarded with remote connectors pulling from my claude.ai account—something I never opted into. I searched for any announcement or setting to turn it off, but found nothing. Disabling it in one session didn’t stick globally, and the UI just keeps listing these mysterious connectors. The whole experience felt invasive and frustrating, leaving me wondering if anyone else has dealt with this and how to completely remove it.

I provided the full path to a script, and claude wanted to search through the directory with find command

I spent an hour watching Claude loop over the same bug, getting nowhere and feeling like I was stuck in a definition of insanity. I finally tossed in a quirky Einstein‑style prompt, and wow—suddenly it fixed the issue on the spot. The sudden turnaround was a relief, turning my frustration into a brief, satisfying win.

I tried using Claude’s summarization in VS Code and was constantly let down—the chat got compacted and almost all context vanished. Every time I returned to a basic task, the tool couldn’t pick up where I left off, making it feel like I was starting from scratch. Coming from Cursor, where summarization was smooth, this loss of context was downright frustrating.