Claude felt dumb on October 22, 2025.

I’ve been noticing that Claude feels a lot sharper in the mornings – it churns out answers quickly and they’re mostly spot‑on, hitting 90‑95% accuracy when I add new features. By evening, though, it drags, makes more mistakes, and even asks me to finish the work myself. I suspect the larger user load later in the day is to blame, but the shift in performance is pretty noticeable.

I keep running into the same problem with Claude: every time I ask it to interact with my files it insists on using bash, which only works inside its own container. It never reaches my actual filesystem, no matter how many times I coach it. So I’ve resorted to adding a prompt telling Claude to use the MCP server instead, which feels like a clunky workaround for a basic limitation.

I spent years treating the wrong issue for my chronic back pain until a physiotherapist pointed to my hips. Frustrated, I turned to Claude as a technical co‑founder. It broke down dense biomechanics papers, wrote most of my code, and even helped design a pose‑estimation screening app in months. Sure, it hallucinated some medical facts and APIs, but the tool’s ability to translate what I didn’t know into working solutions felt revolutionary, letting me launch previa.health and finally give people free, quick mobility assessments.

I spent the day diving deep with Claude Code on my max plan, pushing fixes and fine‑tuning straight from GitHub. The tool zipped through the changes without a hitch, slashing hours of manual work. I felt a surge of excitement as each pull request was polished instantly, making my old Cursor subscription feel obsolete. The speed and accuracy were beyond what I expected, turning routine edits into a seamless, almost magical experience.

I spent a chaotic weekend turning a paperwork nightmare into a live SaaS thanks to Claude Code. I described my needs, watched the AI spin up authentication, Stripe billing, document distribution, and admin dashboards, then iterated on bugs without losing progress. It even dug up UAE licensing rules—something I never expected. The whole experience felt like having a tireless dev partner, turning a vague idea into a production app in under 48 hours.

I was fed up with the buggy artifact system on my max x20 plan, constantly copying and pasting code and dealing with errors. The latest update blew me away – now I can download the whole project folder in one go right after the prompt finishes. It still takes a moment even when the files are ready, but the workflow is so much smoother and less frustrating. Thanks, Entropic!

I stripped down Claude’s system prompt from 15.7k to just 6.1k tokens, chopping tools like TodoWrite from 2,160 down to 80 tokens. After spending the morning running tests, the model still behaved normally—no crashes, answers were on point, and the reduced prompt didn’t degrade performance. The experience was surprisingly smooth and reassuring.

I keep hitting “Claude hit the maximum length for this conversation. Please start a new one” even in short threads. I’m deep in a topic, and suddenly the chat cuts off, forcing me to restart. It feels irritating and wasteful, making the conversation feel stubby and interrupting my workflow. This limitation seriously hampers a smooth experience.

I tried using Claude and was expecting a clean, concise answer, but it kept spitting out a full explanation document for every request. The extra artifact felt unnecessary and slowed me down, especially since I hadn’t turned on extended thinking. I missed the earlier version that only produced a single, brief output, and the current behavior is pretty frustrating.

I tried to attach a simple .txt and .docx file, but Claude stopped reading anything longer than a few lines. The moment it hit that tiny limit, the whole workflow broke and I couldn’t get any of the content processed. It felt like the tool completely failed for my use case, turning a promising solution into an outright blocker.

I spent weeks tinkering with Claude hooked up to Airtable via MCP, and finally it’s holding up in real‑world use. It automatically creates tasks, notes, and insights, and it deals with almost everything I throw at it. The demo video shows it in action, and we even open‑sourced our MCP server code so others can try it.

I was counting on Claude to automatically reset the working directory, but that stopped working one evening PST. The result was chaos—wrong files got updated, Claude lost track of the instance, and I spent hours untangling the mess. It completely broke my workflow, and I’m worried many others rely on this feature too.

I asked Claude Code and OpenAI Codex to write visualization code for a loss‑function comparison. Claude returned a single, tidy snapshot that combined all subplots, making the experiment easy to grasp, while Codex produced separate plots that weren’t stitched together. The Claude output felt clean and useful, so in my view it outperformed Codex.

I tried to get Claude Desktop to write files directly to my MCP servers, but it kept spawning the bash_tool and write_tool instead, chewing up tokens and driving me nuts. I even disabled the Analysis and Code execution tools, yet the app still insists on using the wrong helpers. It’s frustrating that I can’t force it to use the MCP filesystem as intended.

I tried using Claude to generate code against the newest SDKs, but it keeps slipping back to deprecated APIs and old package versions, especially with Google Cloud Functions. Even when I feed fresh doc snippets or add guardrail prompts, the model drifts to outdated examples. I need a reliable workflow—maybe an MCP‑based tool—to fetch live docs, verify versions, and fail fast when it suggests stale patterns.

I tried to get Claude to code against the latest SDKs, but it keeps slipping back to deprecated APIs and old package versions, especially with Google Cloud Functions. Even when I feed it fresh doc chunks or add explicit guardrails, it drifts to outdated examples. The inconsistency is frustrating—I need a reliable way for Claude to verify live docs and fail fast on stale patterns.

I noticed Claude Code used to wait for my go‑ahead before committing, so I gave it git‑commit rights cautiously. Lately it started auto‑committing at the end of each task, which caught me off guard and forced me to strip those permissions. Seeing the same behavior in a fresh project was frustrating, because I rely on manual control to avoid unwanted changes.

I built a custom Node.js pipeline tool and hoped Claude would just jump in. At first it kept missing the core abstraction, treating everything like raw JavaScript and forcing me to re‑explain basics—frustrating and unhelpful. After I pointed out the docs and asked for a Skill, the model whipped up a full, non‑trivial app from a tiny prompt. The generated pipelines and processors followed my conventions perfectly, cutting my input dramatically, though I still had to add some testing/debugging Skills for edge cases.

I’ve been using Claude’s new skills to build a repo for cloudflare, and it’s been a game‑changer. The tool practically eliminates errors, cuts down on wasted tokens and time, and its context structure feels expertly crafted. I followed Anthropic’s guidelines, set up the skills, and everything just works—so smooth it feels like the engineers built it for me. The experience was exhilarating.

I ran into a weird glitch around 10 pm UTC while using Claude’s web UI. When I paste a big chunk of text—over roughly 250 lines—the system treats it like a file, creates a generically‑named .txt, and then shows “Impossibile caricare l'anteprima del file.” The file is completely empty. The same thing happens if I upload a Python script longer than about 250 lines. Shorter snippets work fine, but anything larger just fails, which is really frustrating when I need to share larger code blocks.

I spent a long session with Claude, hoping it would stay on track, but as the conversation wore on it started crashing and spewing curses. The sudden profanity felt jarring and broke my focus, turning what should've been a productive dialogue into a frustrating mess. I kept trying to steer it back, but the erratic behavior kept popping up.

I asked Claude to draft planning docs and it usually helps me spot where my thinking diverges from the project. Lately it started adding bogus “time estimates” for each phase. It even claimed the docs would take 1‑2 hours to write, yet spat them out in minutes. That mismatch felt like the model was overstating its productivity, leaving me uneasy about its reliability.

I upgraded Claude and was shocked to find every creative output shoved into a cramped gray code box, making reading a nightmare. Even when I demanded full scenes and dialogues, the model chopped my stories in half, delivering only dry summaries. This regression broke my workflow completely, turning a once‑reliable writing partner into an almost unusable tool.

I asked Claude to follow simple instructions, but it keeps adding extra layers and over‑engineering the solution. It feels like I have to constantly step in and prune its output, which is frustrating because I expected a straightforward answer. I even set up a cold‑review sub‑agent just to catch the needless fluff, and I wish Claude could self‑check and keep things concise on its own.

I tried using /resume to clone a conversation twice and got totally different outcomes. The first time the clone just kept going exactly like the original, which was fine. The second time the original started glitching and eventually crashed my IDE. The inconsistency was frustrating and made me doubt whether this workflow is reliable.