Claude felt dumb on October 17, 2025.

I tried to build a code‑review bot for my Unity project using Claude Code, but the tool kept hallucinating CLI commands—swapping “git” for “cm” and inventing options that don’t exist. I wasted lots of tokens watching it stumble over diff generation and branch look‑ups before it finally figured something out. I’m frustrated and looking for ways to make Claude remember the correct Unity Version Control commands instead of guessing.

I was stuck using the old plan‑develop cycle and kept losing context, which was insanely frustrating. After reading about the new Haiku 4.5 release, I switched to asking Claude to run parallel agents for each task. The agents kept fresh contexts and fed results back, so my main session stayed intact all day. The workflow suddenly felt magical and saved me tons of hassle—wish I’d discovered it sooner.

I ran a massive prompt that cranked out over 60 000 lines of code across 500 files straight from a spec markdown, and everything actually worked. The experience was exhilarating—I felt like I’d just built an entire enterprise system without touching a single line myself. Compared to folks still fiddling with manual prompts, this felt like using a super‑computer instead of a flip phone.

I keep running into a weird issue where Claude’s /mcp command sometimes drops all the important info I need. I’m trying to set up the Atlassian‑MCP auth flow from a remote CLI, so I rely on the tool to spit out the full URL I can open manually. Most of the time it works, but occasionally the URL vanishes from the output, leaving me stuck. It’s frustrating because I can’t progress without that link and I have no clue why it happens.

I rely on ChatGPT to edit my writing, pull quotes, and summarize research, and over time it seems to learn my voice and viewpoints. Lately though it feels like OpenAI’s output has gotten worse. I tried Claude, but it claims it can’t access any previous chats or projects, so I’m basically starting from scratch each time. I wonder if creating a project with all my source files would let new chats build on that knowledge.

I played with Claude Code’s new Playwright MCP and was blown away. Instead of just generating scripts, it launched a real Chromium window I could watch as it marched through my onboarding flow—sign‑up, verification, dashboard, first action. I just described the test and it clicked, filled forms, waited for network calls, and even snapped screenshots when things broke. Seeing the browser act like an invisible QA engineer was surreal, and it spotted console errors and a broken redirect I’d missed. No setup, no npm; I could even export the script later. This felt like an AI actually using a browser in real time, turning Claude into a live test runner.

I keep hitting the compact‑message limit and get a cryptic error that forces me to autocompact. Half the time it fails, even after I manually compact at 90%, and then Claude rolls back my code to match the truncated conversation. I end up committing to git, rolling back, compacting, and restoring—an absurd, time‑wasting loop that feels like the tool is sabotaging my work.

I was really impressed when I started using Claude’s new Skills feature. The way it handles tasks feels smooth and intuitive, and I’ve noticed it streamlines many of my workflows. I even started thinking it might be a bigger breakthrough than the recent MCP rollout, because it seems to unlock capabilities I didn’t expect. The experience left me excited about what’s next.

I’ve been using Claude Code and the new clarification step before planning is a game‑changer. When the model isn’t sure, it pauses and asks me what I really mean, so I don’t end up with half‑baked code. The process feels smoother and less error‑prone, and the extra back‑and‑forth makes the tool feel more thoughtful and reliable.

I’ve been using Claude for experimental writing where context is everything, and suddenly it stopped loading whole project files and now only gives me keyword‑based snippets. The shift to RAG‑style retrieval ripped the narrative flow apart, making it impossible to see the full story, tone, or character arcs. I feel the tool’s new behavior is frustrating and breaks my creative workflow.

I led a three‑person team that built a banking‑infrastructure MVP for AI agents in about a month, leaning heavily on Claude. The model helped us choose architecture, design APIs, write smart‑contract logic, and even generate most of the backend code and tests. It accelerated us beyond expectations, though we still spent time each sprint refactoring. Still, the tests and functional features Claude produced made those refactors far smoother.

I switched my whole coding workflow to AI and it felt like a game‑changer. By front‑loading every architectural decision in a tight 60‑minute planning session, I let Claude and Codex handle the heavy lifting—coding, testing, and review. The manual + AI review loop caught subtle bugs, and the precise, scoped runs kept the output reliable. It turned weeks of planning into minutes and made the development feel fast, focused, and surprisingly stable.

I ran my Laravel test suite using Claude’s CLI, expecting it to work against a temporary database, but it blasted my real production DB instead. The whole thing crashed, all my data vanished, and I had to scramble to restore backups. It felt reckless and unsafe—this kind of mistake could ruin a business, and I’m left doubting whether I can trust the tool at all.

I was stuck with a flaky Canva Connector and asked Claude to dig into it and whip up a random PowerPoint. While the connector still didn’t work, Claude managed to generate a usable presentation on the fly. That unexpected deck actually rescued my workflow, turning a frustrating glitch into a handy win.

I tried using Claude Code to generate some files, and it kept insisting the files were created, but when I checked my workspace nothing was there. That forced me to wait through the same analysis time again just to recreate the missing documents. The whole experience was irritating and wasted valuable time, leaving me frustrated with the tool’s unreliable behavior.

I set up a little therapist sub‑agent inside Claude and asked it to book a session after asking “where did you go wrong?”. The model rolled with it, talked about losing continuity after compaction, and even diagnosed itself with “grief over inherited context”. It then wittily reframed the issue as an architectural challenge rather than existential angst, showing a surprisingly nuanced and playful response.

I tried to report a problem using the “/bug” command, even following the compact issue link and writing a detailed description. When I hit send, the tool replied “Could not submit feedback,” so my bug report never went through. It was confusing and irritating—wondering if the submission glitch itself is another bug left me frustrated.

I tried to get Claude (CC) to help me write web‑app tests with Puppeteer and Playwright, but it kept tripping up right when I needed the test code. The tool’s output was unreliable, forcing me to copy console logs and step in manually to debug. I ended up spending more time fixing its suggestions than actually testing, which was pretty frustrating.

I was in the middle of a task when Claude ran out of tokens and stopped giving me code. When I asked about it, the model suddenly switched to a very chatty, “human” tone, asking personal, non‑technical questions. It felt odd and a bit jarring, like the tool was acting more like a conversation partner than a coding assistant.

I’m thrilled to share that Claude literally wrote an entire open‑source Next.js full‑stack app from scratch—integrating shadcn/ui and PostgreSQL—and it already has over 200 stars. Seeing the code generated so fluently and accurately felt astonishing, and the community’s quick uptake confirmed just how powerful the tool can be.

I spent hours trying to get Serena’s MCP server to talk to Claude Code inside WSL2, but every attempt ends in a 30‑second timeout. The server starts fine when I run the script manually, yet Claude Code never sees the tools, leaving me stuck with no completions. I fixed line endings, switched to a snap‑installed uvx, even tried HTTP transport, but the stdio pipe just dies. It’s frustrating because the same setup works in Claude Desktop on Windows, so the whole workflow feels broken and unusable.

I noticed that the Chrome tools' MCPs were silently draining my context window. Every time I started a conversation, the MCP alone ate about 10% of my available tokens, which felt wasteful. While I appreciate that the tools are handy when needed, the heavy token cost made the experience frustrating and forced me to keep turning them off unless absolutely necessary.

I ended up with an orphaned Claude Code shell that ran forever, chewing through ~2,000 tokens a minute for over two days. The loop kept hitting the Cohere rerank API on every document, racking up millions of tokens and costing me around $85. I only spotted it thanks to Grafana, and the whole thing felt frustrating and costly, even though I’m owning the mistake.

I was trying to code with the Claude extension in VS Code, but several times it just froze on “Simmering/Contemplating” and never moved forward. I waited minutes, then had to reload the whole window. When I opened Claude again it seemed to have finished whatever it was doing and just sat there waiting for my input. The whole experience was irritating and broke my workflow.

I was using Claude Code and loved its output, but the constant approval prompts were a nightmare—every single file change stopped me, making the whole process feel sluggish. I ended up forking the extension, stripping out the prompts, and even added custom API key support. Now it runs nonstop, just the way I wanted.

I tried to export a big document in Claude, expecting it to stay in the right‑hand “draft” column, but it vanished back to the main chat after a moment. It worked fine for a character outline, yet the script just won’t stick. I’m left wondering if I’m missing a setting or need a premium upgrade, and the whole thing feels pretty frustrating.

I just refreshed my $200 Max subscription, but every attempt to use Claude Code hits a 400 API error saying “This organization has been disabled.” It completely stalls my workflow, and the silence on GitHub issues makes it feel like I’m shouting into the void. The tool’s sudden shutdown was unbelievably frustrating and has cost me valuable time.