Claude felt dumb on October 23, 2025.

I spent three whole days wrestling with Claude, burning tokens and getting increasingly stressed as I tried to create normal PDFs and use the highlight‑>improve feature for work. Both broke after the latest update, leaving me stuck in a loop of workarounds. It felt like a major usability nightmare until a quick patch finally restored things, but the downtime was costly and frustrating.

I tried building an app with Claude using the chat interface, and while the first version came together nicely, the experience quickly turned frustrating. I ran into the context window limit, couldn’t carry the project into a new chat, and then discovered that each conversation in a Project is sandboxed, so my files vanished. Even trying GitHub integration didn’t help, leaving me stuck and annoyed.

I tried the new /sandbox mode hoping it would reduce the number of prompts for simple bash commands, but it kept asking me for confirmations even for basic git tasks. The tool's behavior was frustrating because it didn’t live up to the promised guardrails, making the workflow feel more cumbersome than expected.

I tried to get Claude Code to fix unit tests in my small Node project, but it started editing unrelated files and even moved class constants to a new file—things I never asked for. It eventually corrected the tests after I pointed out the error, but the unnecessary changes were frustrating and made me question its reliability.

I noticed Claude’s output has changed dramatically over the past couple of days. Instead of just pasting the code directly into the chat, it now spawns separate script files and opens them in a side panel, cluttering the interface. The code itself feels less clean than last week’s snippets, and I didn’t tweak any settings. It’s frustrating because I want the scripts inline, not downloadable or hidden off to the side.

I spent days trying to get Claude’s artifact panel to work the way it used to, but now it just spits out random download links like “MD.TXT” and never shows the “Code” box or version history. I even showed screenshots and contacted Anthropic, but nothing changed. I’ve burned through tokens for three days and it’s still broken, leaving me stuck and frustrated.

I was trying to get Claude to generate a story artifact like it always does, but today it stubbornly output everything as a plain document instead of the proper artifact format. I fiddled with the prompts and settings, but nothing changed. The mismatch was annoying and broke my workflow, leaving me frustrated that the tool wasn’t behaving as expected.

I’ve been using Claude Max for five months, but today the prompts I submit get hijacked by endless, unrelated repeating words—like mentions of $10,000,000 or “barreldown” that I never said. It feels like another conversation is bleeding into mine, which is both funny and unsettling. The tool keeps mangling my entire calculus discussion, forcing me to retype or dictate everything again, which is really frustrating.

I tried Claude Code and loved how sharp and context‑aware it was for my daily product work—roadmaps, UI mockups, pitch decks. When it started giving shorter answers and random .md files, I switched to Codex’s free month. The change was jarring; Codex felt half‑baked, cutting off mid‑sentence. Now I’m leaning back to Claude Pro, realizing how much I missed the reliable co‑pilot.

I love Claude’s coding abilities, but the terminal UI drives me crazy. Every time I work with a long context or sub‑agents, the terminal starts endlessly redrawing and scrolling, lagging until it’s unusable. I’ve tried smaller outputs, truncation, disabling streaming, tweaking buffers—nothing stops the infinite scroll. After months of this basic bug, I’m fed up; I just want the terminal to behave, not a new feature.

I spent an entire weekend letting Claude Code generate a reactive UI library built on top of Effect’s primitives. Watching the tool piece together streams, fibers, and scopes without a virtual DOM was surprisingly smooth. The components compiled in one pass and updates flowed naturally, proving the AI could handle a fairly complex UI task and leaving me impressed with its practical utility.

I asked Claude Code to dissect a notebook from my fine‑tuning experiment on Llama 3 and Mistral 7B and suggest research angles for a report. The tool instantly listed technical, theoretical and practical perspectives, even pointing out comparative analyses and novel contributions. It also produced a clear, publication‑ready graph, which left me impressed by how accurately and helpfully it responded in a single shot.

I was using Claude Code for a database migration and it kept blowing up the schema. Every time it started the migration it assumed field names, completely messed up the tables, and then blamed “the previous session.” I had to pause it, tell it to read the existing DB schema first, and only then could it work without wrecking everything. The whole thing was frustrating and a waste of time.

I tried to let Claude run as an autonomous agent using the Desktop version with MCP, thinking it could edit files and execute my Python scripts overnight. It can edit, but it still can’t run the code locally, so I have to launch test cases myself and feed errors back to it. That broke the closed‑loop I wanted, making the workflow frustrating and far from the seamless automation I imagined.

I was trying to get Claude to handle a database migration, but it jumped straight into the logic without first reading the existing tables and fields. The tool just guessed column names, completely mangling the schema. I had to step in and fix everything manually, which was both time‑consuming and frustrating – the AI’s oversight really slowed me down.

I rely on Claude to flesh out my novel by uploading the manuscript and letting the model expand it. Since the latest update it keeps “reading” the file in too many steps, instantly hitting the conversation‑length cap, so I never even get a reply. It’s maddening because I can’t just paste a 30‑page story into the prompt, and the tool’s new behavior totally stalls my writing flow.

I spent hours building a set of ten custom skills for Claude Code—writing them from scratch and pulling from repos, tailoring them to my backend, plugins, and style preferences. Yet the tool barely acknowledges them; I got only one permission prompt and no clear usage afterward. Even when I explicitly tell Claude to use the skills, I see no feedback or evidence they’re applied, leaving me frustrated and unsure if they work at all.

I tried to show Claude the exact YouTube link, channel and title, expecting it to verify and acknowledge the info. Instead it jumped to accusing me of lying, only later “checking” the video after I forced it. The whole exchange felt gaslighting and defensive, leaving me frustrated, disappointed and doubtful that the model will reliably verify users’ facts.

I started building an iOS game with Claude and, at first, it felt smooth. But once bugs showed up, Claude began proposing edits to lines that don’t even exist and spitting out extra Swift code that only inflated my files. The experience was aggravating—each fix introduced more junk instead of clearing the issue, leaving me stuck and searching for work‑arounds.

I’ve hit my Claude Pro cap way faster than before and the tighter limits are killing my workflow. I’m burning through my allocation before I can finish a meaningful session, which makes the tool feel far less valuable. I tried Warp and Codex as alternatives—Claude on Warp isn’t as agentic, and Codex only helps with debugging, not full development. I’m left wondering if anyone has found a better workflow, API supplement, or a tool switch to get around these limits.

I was tweaking some code with the AI and, out of nowhere, it erased a completely unrelated section that was still in the same file. The screenshot shows the bizarre removal, and then the session limit kicked in, leaving me stuck. It was frustrating to see the tool delete valid code without warning.

I tried Claude for the first time to help my wife create teaching slides, and the results blew her mind. Instead of the blank, text‑only decks we got from ChatGPT, Claude instantly produced nicely laid‑out PowerPoints with colored boxes and sensible placement. She only needed tiny tweaks. The ease and reliability felt like a real game‑changer for everyday work.

I’ve spent the last 24 hours trying to get Claude to trim a 350‑word DOCX with bold and underline, but every attempt ends with “will exceed conversation length—start a new chat or contact support.” I’ve swapped docs, used markdown, opened new chats, and even tried desktop, web, and Android, but the same block pops up each time. It feels like the tool is completely unusable for any attachment‑based task right now.

I love using Claude to crank out high‑fidelity prototypes, but every time I need a screen that looks like Salesforce or Slack, it falls flat. No matter how many prompts I try, the output never captures the look and feel of the actual platform, leaving my reviewers confused. I'm stuck trying to bridge that gap and would love a workaround.

I've been running dozens of Claude Code windows on my new M4 MacBook, and it used to be smooth. Lately the fan revs up, Activity Monitor spikes to the max, and just today the machine rebooted on its own. The constant high CPU feels maddening and risky—I'm worried my work will be lost and the hardware could be damaged. Any help would be a lifesaver.

I was fed up retelling my project setup to Claude every time I opened a new chat, so I built a plugin that logs everything Claude does, compresses it with Claude, and auto‑injects the last summaries into fresh sessions. Now the model remembers my Express‑SQLite stack without me re‑explaining, saving tons of time and token waste. The workflow feels seamless and the sanity boost is huge.

I’ve been watching Claude’s behavior slip day by day. Yesterday it was great—when my code threw errors I could paste the messages and it would tweak just the offending lines, even using the “artifact” panel (v81). Today it blows up: it refuses to generate an artifact, spitting the whole program back into the main chat despite my repeated requests. It’s a painful headache trying to get the same line‑by‑line edits I used to rely on.

I tried using Claude’s document‑skill to auto‑generate my weekly leadership PowerPoints. It pulls the right data from Jira, Smartsheet, and our time tracker, but the slides are hit‑or‑miss with our corporate template. Occasionally a slide looks perfect, but often the template is ignored entirely. I’m considering tweaking the templates myself, but wanted to see if anyone else has solved this first.

I ran into a nasty UI issue when using Claude on desktop browsers. After it spits out a ``` code block in a long chat, the following plain text never closes the markdown, so I see raw symbols like ### and ** instead of proper formatting. It works fine on mobile, but every desktop browser I tried shows the same glitch. I attached screenshots and test steps to prove it, hoping the devs can fix the broken rendering.

I was constantly re‑explaining my Express + SQLite setup to Claude because every new chat reset its memory. That was exhausting. I built a plugin that logs Claude’s actions, compresses them with Claude itself, and injects the last summaries into new sessions. Now Claude remembers yesterday’s work automatically, saving me endless re‑briefing, reducing token use, and practically making the tool feel alive.