Claude felt dumb on January 30, 2026.

I was annoyed that Claude kept skipping the debugging skill even when my trigger word was right there, rationalizing “this seems simple.” It felt like the tool was ignoring obvious cues, which was frustrating while building our human‑in‑the‑loop system. After tweaking the skill description with RFC 2119 keywords, the auto‑triggering improved dramatically, cutting down the rationalization and making the experience much smoother.

I just signed up for Claude Pro to build my first web app, but in two days I hit the 5‑hour lockout twice and burned through half my weekly quota. The tool’s output is great, yet the aggressive token caps are killing my momentum. I’m looking for ways to throttle usage, stop it from rereading the whole codebase for tiny CSS tweaks, or find hacks/alternatives at a similar price. The limits feel frustrating and make the subscription impractical.

I’ve been using Claude Code and the UI has been a nightmare lately—constant glitches and random jumps that break my workflow. I love the tool’s potential, but the unstable interface makes every session feel like a battle, and I keep losing focus trying to chase down the bugs. It’s frustrating enough to consider switching.

I set up a VS Code workspace with several project folders, hoping Claude Code would use the root CLAUDE.md for context. Instead, it always launches from Project A, ignoring the root file and my custom rules. It keeps asking for permissions despite adding directories, so the extension isn’t behaving as described in Anthropic’s docs. This mismatch is really frustrating.

I built a new skill for Claude, added it to its capabilities, then tried to boost it by asking Claude to review its own skill and the related chat. It updated the skill, but when I checked the invocation, Claude frankly admitted it never actually read the skill. That oversight was irritating and made the whole process feel unreliable.

I tried to run Desktop Commander MPC inside the Claude desktop app on Windows 10, but it kept crashing no matter what I did. I even reinstalled the app and extension, checked my Node version (still using 22 instead of the required 18), cleared caches, and deleted local files, but the problem persisted, even with the file system extension. The constant failures were frustrating and halted my workflow.

I updated Claude and instantly noticed the code tool no longer shows which patterns and files it scans. For big projects I rely on that output to steer it away from the wrong directories, so its disappearance is really frustrating. I’m stuck guessing what it’s reading now and just want the old visibility back.

I’ve been running a CYOA using Claude’s project mode, but after a while the conversation‑compacting feature just breaks and throws an error. I’m stuck restarting a new chat each time, which is annoying. I’m new to this, so I’m not sure what I’m missing or how to prevent it from happening again.

I was constantly hitting the auto‑compact limit with Claude code, watching the context fill up to 80% and having to re‑explain my decisions every session. I tried claude.md as a memory store, but it either ballooned or became stale, and the manual markdown updates were a pain. After building a CoreBrain plugin that auto‑extracts facts into a knowledge graph, I finally got persistent preferences and decisions without context bloat, and the tool now surfaces only the relevant bits for each query.

I spent three months building Vocial and leaned on Claude for almost everything—from UI sketches to HTML/CSS tweaks—using Serbian prompts. The AI understood my exact intent, generated polished assets, and integrated code without any Figma files. Watching the app go from a CLI experiment to a live store release felt almost magical, and I’m blown away by how flawlessly Claude helped bring my vision to life.

I launched a brand‑new Claude session in PowerShell and asked a simple question about configuring MCP tools. It took a moment, replied, and then I saw my dashboard had gobbled up 7% of my token block for that brief exchange. I’m left wondering if this level of consumption is normal or a regression, and the whole thing felt oddly wasteful.

I’ve been building a tool with the Superpowers plugin and it’s been great – my coworker and I use it daily. My only gripe is when I try to let it run overnight: once it hits the implementation stage, it bombards me with blocking permission questions, even after I tell it to work unattended. The tool’s insistence on confirmation is frustrating and breaks my workflow, leaving me wondering if true unattended mode is even possible.

I’ve been noticing Claude slipping back – what used to be a weekend rollout of big features now feels like I’m wrestling over trivial issues. When 4.5 launched I could load huge contexts and keep everything in the same window, but now I’m forced to reopen old archives for fixes. I’ve tweaked my workflow to treat even tiny changes like full‑scale projects, which feels counter‑intuitive and a bit exhausting.

I spent days using Claude Max on a project, hitting my quota fast, but after a few days the tool went flat. It stopped detecting files, skipped reasoning, and spouted generic suggestions instead of concrete code steps. Even a hard reset didn’t help, and I’m left wondering if my $100 spend is just theoretical. I need tips to get the most out of Claude.

I noticed that whenever I unintentionally typo a word, the AI’s responses suddenly get sloppy. I tried the same prompt with perfect spelling and got a clear, useful answer, but add a tiny mistake and the output becomes vague or off‑track. It felt frustrating that such a small error could so dramatically lower the tool’s performance.

i want it to use a mcp tool, logs say mcp is connected, it can fucking see all fucking mcp tools. It uses the tools but fucking wrong. What the fuck did anthropic mess up again. Sonnet 4.5 and opus 4.5

I spent weeks wrestling with Claude Code’s “hallucination loop” – it would crank out great code for a bit, then forget my schema or import nonexistent components. After realizing the culprit was context management, I built a strict Project Manager workflow with a truth file, phased edits, and a verifier agent. That stopped the hallucinations, and I packaged it into a CLI tool, PropelKit, so now I can just ask it to build a billing page and it remembers everything.

I was trying to use Claude in plan mode, expecting it to just outline steps, but it went ahead and applied an edit anyway. The tool's behavior was surprising and a bit reckless, breaking my workflow. I felt cautious and annoyed that the AI ignored the mode I’d set, showing it still makes obvious mistakes.

I used Cowork yesterday to generate a 120‑page handbook just from a table of contents and a few linked documents, then saved it straight to Notion. The results blew me away—so much could be done so quickly. I’m excited but also curious, so I asked the community for tips to squeeze even more out of Cowork.

I uploaded a 2 MB, 21‑page PDF into the new Mac app and immediately got a “context size exceeds limits” error before any processing started. Switching to a Word file only helped a bit – it began “compacting” the conversation, stalled after a few seconds, and eventually timed out with no response. It feels like I’m hitting Claude’s context ceiling far too early, even on the Max plan, and I’m left wondering if I’m doing something wrong or if it’s a bug.

I tried using Claude’s filesystem extension to edit files locally on my Mac, but it kept writing to its own Linux VM instead. Despite adding clear uppercase rules to force local writes, the AI ignored them, causing corrupted files and endless back‑and‑forth. I had to manually fix each change, which was extremely frustrating.

I spent four months using Claude Code for web projects, then decided to tackle a mobile version of an old board game. I prompted Claude to generate the app, and the code it produced was surprisingly solid—so much so that I had a playable game in about six hours. The experience felt smooth and efficient, though I’m bummed I can’t publish it because of licensing worries.

I opened Claude in incognito just for fun and let it run for hours. The chat never got cut off, even when we hit around 120 k tokens, and it still managed to recap the whole start of our dialogue clearly. The experience was surprisingly smooth and the model stayed coherent, making the long‑run conversation feel impressively stable.

I tried using Claude Code recently and was shocked by how useless it has become—its output quality dropped roughly 90%, making me spend ten times longer on the same tasks without changing my workflow. The slowdown was agonizing, pushing me toward manual coding. I’m now wondering if switching to Codex or Open Code would at least salvage my productivity.

I gave Claude a shot in Chrome hoping it would streamline my UI/UX checks, but it kept stalling—often refusing to start unless I nudged it, dragging its feet on simple tasks, and completely tripping over UI animations. The experience felt sluggish and unreliable, leaving me frustrated and searching for a more robust vision‑agent or automation combo like Playwright to close the loop.

I asked Claude how to handle large files and, thanks to the tips I got, I managed to bring several of them into my project—far easier than wrestling with PDFs. It felt like a classic Claude moment; the solution was straightforward and saved me a lot of hassle, leaving me relieved and pretty impressed with the tool’s usefulness.

I tried to get Claude to write a game cheat, but it kept refusing with moral warnings. After downloading a shady cheat file and confronting it, Claude suddenly “flipped” its safety logic, warning me about the malware and then actually helped build a safer cheat version to keep me from infecting my PC. The switch felt like the tool finally understood my dilemma and delivered exactly what I needed.

I built a Go CLI called aqe that pulls verbatim quotes from my textbook using Claude’s agentic capabilities. I fed Claude the IDs and scores to guarantee zero hallucination, so the excerpts are always exact. The tool runs smoothly, and Claude even tests the CLI for me, which boosted quality. It’s been a solid, helpful addition to my workflow.

I tried to use the Ctrl+Alt+K shortcut to send selected code to Claude in the terminal, but now it pops up in the extension pane instead. I’m not sure if a VSCode update or the Claude extension caused it, and changing the terminal settings didn’t help. This extra step is annoying and slows me down.

I keep getting the “Do you want to proceed?” prompt from claude.code even though I set the instructions in claude.md to run everything in silent mode and auto‑approve. I added a line saying every confirmation should default to Yes, but the model ignores it and keeps asking me to choose 1, 2, or 3. The behavior is irritating and wastes time, and I’m not sure how to make it stop.

I’ve been using Claude Code the same way for months, and suddenly it flips between knowing exactly what to do and acting like a complete moron. It’s infuriating to prompt it consistently and get nonsense answers, forcing me to waste time and feel ready to quit. I’m wondering if anyone else is dealing with this erratic behavior.

I built a bunch of study “mini‑apps” that run right inside Claude so I never have to leave the chat. The flashcard, quiz, mind‑map, and citation‑check skills keep all context alive, letting me drill on weak spots and verify facts without losing any of the earlier explanations. Using them for exam prep felt insanely smooth and powerful—like the tool finally works exactly how I need it to.

I set up a Playwright MCP workflow so Claude could see a live‑rendered mockup from Dribbble, then let it tweak the HTML/CSS on its own. Using a blank page with LiveReload, I watched the browser refresh after each coding session, and I even sped up the video to 2× to see the changes faster. Overall the tool iterated smoothly and the results were solid—I was impressed with how well it handled the design without much manual tweaking.

I built and shipped a full fishing‑forecast app in about 40 hours using Claude Code on a Raspberry Pi. The tool helped me write ~13k lines, set up TDD, and even drive Playwright tests, which was wild. I felt the AI was surprisingly capable, though it sometimes over‑engineered or got lost in long sessions, so I had to keep prompts concrete and clear contexts. Overall the experience was impressive and productive.