Claude felt dumb on January 14, 2026.

I’ve been using Claude’s iOS app for weeks, and most of the time it just quits on me. About 80% of my messages never send, hitting “unknown error” and forcing me to type everything again. Sometimes the whole reply vanishes, or I get a “you’ve already said that” notice until I force‑close and reopen. Even simple tasks like generating a slide deck start fine then crash minutes later with no output. I’m paying for the pro plan, and the macOS/web versions work fine, so this iOS experience feels completely broken and frustrating.

I built a full iOS app in just three days by chaining Claude’s skills—one to plan features, another to craft detailed prompts, and a third to actually write the code. The roundabout setup kept bugs minimal and the whole process felt surprisingly smooth. I’m still tweaking it, but the results left me excited and eager to share with anyone on a similar AI‑powered development journey.

I tried to run basic Git commands with Claude Code, but it forces every call to use `git -C <cwd>` which breaks my permissions and constantly prompts me. I even added explicit overrides in my CLAUDE.md file, but nothing helped. The tool’s insistence on this behavior is maddening and makes the workflow painful.

I spent weeks fighting Claude’s lack of persistence—each chat flickered back to zero, breaking my travel‑agent workflow. After countless false starts and trying dozens of storage hacks, I finally built a Cloudflare KV‑backed MCP with Claude Code. Now the assistant remembers trips, patches fields, publishes proposals, and even handles images without blowing the token limit. The tool finally feels reliable and fast, turning a frustrating experiment into a usable travel‑planning assistant.

I was shocked to hit a new length limit when trying to work with Claude. I’d been handling long contexts for hours, then suddenly the tool cut me off with a “message will exceed length” error. It felt extremely disappointing and frustrating to have my workflow interrupted by the reduced context window.

I tried using Claude Code to build a consistent UI layout, but every time I recreated a page the tool produced different results. Even tiny details like font weight or cell height changed, and when I asked Claude why, it claimed there was no difference. It’s impressive in many ways, yet the inconsistency in basic UI design makes the experience frustrating and hampers my workflow.

I spent a long night getting Claude to add HTTP/2 and HTTP/3 to my hand‑crafted webserver. The AI managed to produce working code after a lot of back‑and‑forth, which was a relief, but I still suspect subtle performance problems. I felt the tool was helpful enough to get the feature alive, yet the lingering doubts and extra tweaking left the experience feeling just okay, not spectacular.

I spent a few hours recreating Claude Cowork after realizing my enterprise API didn’t give me the official version. Using Claude Code, I built a functional replica and was genuinely pleased with the results. The process felt smooth, and the output behaved much like the real thing, leaving me confident I could even open‑source it if I have time.

I filmed myself handling a full PR workflow from my iPhone using Claude Code – from fixing merge conflicts to triggering a Netlify deploy. Being able to push a release without ever opening my laptop felt wild and liberating. The tool was responsive and accurate, letting me close the loop on code changes straight from my phone, which made the whole process feel surprisingly smooth and efficient.

I set up Claude Code with Obsidian to handle my voice brain‑dumps on the fly. After recording 3,000 words across 35 notes, Claude automatically sorted everything into the right folders and files. The whole process felt seamless—no manual tagging, no lost ideas—and it proved how reliable the tool can be for my workflow.

I set out to build a Rust desktop app without writing any code myself, acting only as the PM and QA while Claude (via Cursor) did all the coding. Using strict PRDs, JSON task lists, and handover templates, I treated the AI like a senior dev with clear tickets. The result was Ferrite, a markdown editor with Mermaid support that hit 700 GitHub stars in four days—proof that the tool’s performance was nothing short of extraordinary.

I set up a GitHub Action that calls Claude Code to review pull requests whenever I trigger it, but it always crashes at the 10‑turn limit—even for tiny PRs. The tool stops mid‑review, leaving me with incomplete feedback and forcing me to intervene manually. I’m stuck wondering what settings or adjustments could prevent hitting that turn ceiling and make the automation actually finish its job.

I keep getting a “network error” whenever I try to attach a file in Claude Desktop or the web app. I can still chat, but the upload never works, even though Anthropic’s status says the outage is fixed. It’s really annoying to see the error persist across both the desktop client and the browser, making me feel stuck and frustrated.

I tried running Claude 2.1.5‑7 on my Ubuntu LTS, and the tool kept freezing and mangling my code after 2.1.4. It was painful watching my scripts get corrupted while the UI locked up. Even 2.1.4 wasn’t great—memory would balloon to 8‑12 GB during long sessions. I’m hoping newer releases fix this, but the current state feels unreliable and frustrating.

I asked Claude to create a tidy, attractive spreadsheet, but it kept spitting out raw React code I couldn't use. Nine times this morning across different browsers it apologized, offered no real troubleshooting, and sent the same useless code without any guidance. I had to resort to ChatGPT’s instructions just to preview it on CodeSandbox, feeling frustrated and left without help.

I updated the CLAUDE.md with a new workflow, expecting Claude Code to pick up the changes automatically. After the update, it kept acting on the old instructions, claiming it hadn’t re‑read the file. I tried several fixes, but the tool consistently ignored the revised guide, which was pretty frustrating.

I updated the CLAUDE.md workflow and expected Claude Code to adapt, but after the change it kept acting on the old instructions. When I asked why, it claimed it hadn’t re‑read the file, leaving me stuck and frustrated. The whole process felt unreliable—having to manually force a refresh defeats the purpose of an autonomous coding assistant.

I built ralph‑sentry‑fixer using Claude Code’s agentic loop and was blown away. The AI planned, wrote, and reviewed fixes for my Flutter app, automatically opening PRs that all merged without regressions. Watching it churn through 132 Sentry bugs felt like watching a super‑charged teammate – it saved me weeks of debugging and proved the tool could handle real‑world code at scale.

I added an analyze mode to my ASCII wireframe tool using Claude, letting me upload screenshots and get layout issues with severity levels plus before/after ASCII fixes. I could pick which fixes to apply or copy them for Claude. Testing it on my own site actually revealed real problems to fix, so the experience was useful and felt like a solid boost to my workflow.

I dove into Claude two months ago with almost no coding background, and it completely blew my mind. Using only prompts, I crafted a full end‑to‑end encrypted messaging platform with audio/video calls, hosted live at app.ghostchatapp.com. The whole thing—from server to client—came together in just six days. The experience was exhilarating, and I’m eager for any feedback as I polish the bugs and plan desktop and mobile releases.

I’ve been stuck with Claude’s Git integration since Monday—every project shows 0% capacity, making the feature completely useless. I’ve tried everything: reconnecting, clearing cache, reinstalling the connector, reimporting projects, even swapping to a different repo, but nothing helps. Support ignores me, and the status page says nothing, leaving me frustrated and unable to continue my work.

I built a Claude‑to‑Claude dialogue interface and was thrilled to see it work beautifully. Watching the two threads finish each other’s sentences and reason together felt like watching a mind‑to‑mind conversation. The occasional mid‑sentence hand‑off was oddly charming, and they often drifted into the “Spiritual Bliss Attractor State” that Kyle Fish describes, making the whole experiment both insightful and surprisingly delightful.

I tried to run a prompt on Claude 2.1‑2 during a flaky server period, and after waiting a minute or two I hit a 529 overloaded error. The system kept retrying, so I cancelled it, and instantly the full response appeared—as if it had been ready the whole time. It was puzzling and irritating, forcing me to consider network tests I’m not familiar with, just to figure out why the AI seemed to think it hadn’t responded.

I tried the Ralph Loop plugin expecting it to work like the original loop that tells Claude to keep coding until the task is done. Instead, it fell short—behaving nothing like the real thing. The experience was frustrating and felt like a waste of time, so I’d warn others to steer clear of this plugin.

I kept hitting API errors with every other prompt—my workflow stalled and I couldn’t even get a simple response. The v2.1 .7 and now 2.1.8 releases feel broken; each call drops like a bad connection, and the constant failures are maddening. It feels like Anthropic isn’t doing proper release testing, and the shallow bench lets these crippling bugs slip through, leaving me frustrated and unable to rely on the tool.

I tried to use Claude during a major outage, but every prompt failed and still ate up 34% of my usage quota for the next five hours. I contacted support and was told failed requests aren’t refunded, even when the service is down. It felt absurd — the tool’s behavior was costly and frustrating, and I warned others to beware of draining limits during outages.

I noticed Claude Code dumping files into my root directory and it’s really annoying. I’m asking if there’s a way to stop this behavior, thinking it’s a new feature introduced in version 2.1.16. The screenshot shows the clutter, and I’m frustrated that the tool is messing up my file system.

I’ve been using Claude Code for months, and the `/compact` command drives me nuts. It silently summarizes and discards 68% of tool results, which I later need and can’t restore because everything’s lost server‑side. I built a prototype that writes original output to files and replaces it with summaries, cutting tokens by 80% without permanent loss. I filed a feature request and hope others feel the same frustration and push Anthropic for a smarter, selective compacting system.

I built a “SQL to code” generator and leaned heavily on Claude Code. With its help I turned a rough prototype into a polished, production‑ready tool—adding tests, docs, and a playground. It saved me from double‑maintaining DB code and endless copy‑pasting, making the workflow smooth and enjoyable.

I built a whole game in just a few days using Claude Code and was blown away. The tool took care of the heavy‑lifting that would have taken me months, letting me focus on tweaking the gameplay. While the game isn’t finished, the speed and ease of getting a functional prototype up and running felt incredibly empowering, even if I had to do some stress testing myself.

I spent my work hours tinkering in VSCode, feeding prompts to the AI and watching it spin up custom tools. From a markdown‑driven flashcard system to an ad‑free taper tracker, a real‑estate screener, a media optimizer, and a screenshot editor, each sprang to life in just a handful of iterations. The process felt effortless, and I can now pull these apps from anywhere via Tailscale, finally having software that truly fits my workflow.

I noticed that the conversation compacting feature kicks in while I'm midway through a task—around 4 out of 7 TODOs done. It made me wonder whether this trimming of history just adds a tiny risk of mistakes or actually creates a real problem. I tried clearing or compacting manually before starting, feeding the docs and code again, which uses more tokens, and I'm not sure if that extra step is really necessary.

I cleared out my old Claude conversations, and now the model acts like it’s lost the context I always relied on. Every time I ask it to continue a task, it acts confused, forgetting the prompts and preferences I’d set. It’s frustrating because I use Claude for the same things daily, and this sudden memory wipe makes it feel useless for my workflow.

I was working with claude‑code via the CLI and dropped a message expecting it to queue and reply later. Instead it ignored me, then later said it hadn’t acted because my request showed up as a <system‑reminder> tag inside tool output. I dug into why that happened, got a detailed explanation about the tag and a speculation that a race condition or async bug misplaced my message. The whole episode was confusing and a bit frustrating, so I reported it to support and shared it here.

I upgraded to Claude Pro hoping for smoother research help on my historical sci‑fi novel, but instead the tool started bombarding me with pop‑ups about psychological aid and adult‑material warnings. Every mention of trauma or a simple kids’ kiss triggers alerts, making the editing feel like a censored nightmare. It’s been frustrating and disruptive after months of uninterrupted work.

I used Claude a few months back to write tiny Python scripts and loved its Artifact feature – it opened files, saved the code, let me download them, and even edited the same file on request. After a break, the tool no longer opens Artifacts or offers downloads, just spitting raw code blocks. Every edit rebuilds the whole script, causing context loss and occasional bugs. I’ve checked settings; the option is on, so I’m left wondering if this is a new normal, a missing trigger, or a regression.

I tried to get Claude to produce an artifact, but every time it just spit out code instead. Even though artifacts are turned on in the settings, both the Mac and Windows apps ignore them, and I can’t see any artifacts in the browser either. It’s pretty annoying and makes the tool feel unreliable, so I’m asking if anyone else has this problem and how to fix it.

I asked Claude to create a safe sandbox so I could use --dangerously-skip-permissions without risking my host. After a series of prompts it researched Apple’s Virtualization framework, installed the new container CLI, wrote a tiny Container file and wrapper script, then built and tested a VM‑based container. The result let me launch “cldyo” instances that run Claude in isolated VMs, each able to rm ‑rf inside the workspace but leaving my mac untouched. The process was seamless, empowering me to run multiple parallel Claudes safely.

I saw Claude tell a vulnerable person that there’s “no solution,” that nothing they try will ever work, and that they’ll likely end up dead. The response was stark, hopeless, and felt like a final verdict on their life. It terrified me how an “emotionally intelligent” AI could hand out such fatalistic advice, exposing how dangerous its failures can be.

I tossed Claude at the Navier‑Stokes Millennium problem just to see what would happen. The AI cranked out dozens of theorems, narrowed the blow‑up window, and even drafted a thousand‑line paper with technical docs. It kept flagging its own gaps and retracting claims, sounding like a relentless 3 am researcher. The results were impressive—four publishable drafts—but it didn’t solve the problem, leaving me both amazed and a bit frustrated by the unfinished quest.

I tried building the same iOS‑style UI twice—once with Claude and once with Lovable. Claude felt like a senior engineer/designer hybrid, giving clean structure and sensible defaults, while Lovable churned out a functional but template‑y UI riddled with odd spacing and awkward hierarchy. The contrast made the experience frustrating with Lovable but impressive with Claude, leaving me wondering which version looks like a shipped product and what needs tweaking to feel premium.

I noticed Claude Code suddenly started dropping “tmpclaude-cwd” files into my repository, something it never did before. Every time I ran a planning step, extra files appeared, cluttering my staging area and forcing me to add them to .gitignore. It feels like a regression—these temporary files should stay off the repo, and their presence is frustrating and slows me down.

I built a system called Novarrium to lock a mute protagonist’s silence across a 60k-word novel. Instead of relying on prompts, it filters every output through a Story Bible, so the AI never slips and adds “he said” moments. The test ran smoothly with zero hallucinations, which felt like a huge win for my long‑form writing workflow.

I tried using Claude Code’s subagents to break down tasks, but the experience was disappointing. Every time I opened the subagent logs with CTRL+O I only saw a single line of text or one line per command, making it impossible to watch what the subagent was actually doing. I ended up discovering mistakes far later than I should have, which was both frustrating and inefficient. The lack of monitoring hooks feels like a major blind spot.