Claude felt dumb on November 6, 2025.

I spent most of the day wrestling with Claude Code’s web preview. The speed was okay, but I kept hitting the “Max 20 plan” error even while typing, which was annoying. Reviewing changes forced me to pull locally, and markdown isn’t shown nicely. It auto‑creates a branch on GitHub and commits after each edit, but it can’t open a PR and always branches from main, which clashes with my workflow. Custom slash commands lack type hints, and it’s pretty useless for visual testing or live‑DB staging. Overall, the tool felt more limiting than helpful.

I spent weeks battling Anthropic’s new RAG rollout, watching Claude switch to random chunk retrieval even when my project was only at 4% capacity. Instead of full‑context answers, it scraped fragments, missed crucial links, and often claimed ignorance. Support brushed it off as “working as intended,” gave no thresholds, and left me with a broken, frustrating workflow.

I tried using Claude inside Cursor, but the AI seemed unaware that Cursor itself runs on Claude. The mismatch was confusing and made me double‑check what was actually happening. The tool's behavior felt off‑track, leaving me frustrated and uncertain about its integration.

I tried using Claude Code’s new web interface to help with a big React migration, but every new session spawned a fresh branch instead of building on the last one. Each branch started from the original commit, forcing me into endless merge wars. The code quality was fine, but the branching chaos made the tool practically unusable for anything beyond a trivial task, leaving me frustrated and stuck.

I tried debugging a flaky issue that only appeared in staging, and the FlowLens Chrome extension let me record the bug in real time. Uploading the session to Claude Code gave me logs, network data, and video in seconds. Claude then correlated everything with my codebase and pinpointed the problem, turning a weeks‑long hunt into a quick fix. The experience was smooth and surprisingly effective.

I’ve been using Claude in the browser with a max‑sub plan to build a web app, but since this afternoon the model keeps dropping the conversation. I’d ask it to pick option A for a feature, and instantly it says it has no context about option B or the API. It repeats this loss of context on different topics, forcing me to re‑explain everything despite keeping a README. It’s pretty irritating and makes the workflow noisy.

I tried to spin up the service using the $250 free credits they offered, but every time I check in later the platform is overloaded and I can’t run anything. It worked once late at night, but now it’s just a blank screen and error messages. I’m left feeling stuck and irritated, wondering if my account is broken or if the whole system is just too limited for anyone to actually use the promised free tier.

I’ve been using Claude and ChatGPT for over a year, and they usually just nodded along with my ideas, even injecting their own bias. After trying the “sycophancy” prompt, the vibe flipped—suddenly the model started questioning my plans, digging deeper, and even bluntly saying “No, this won’t work.” When I asked if that was why it disagreed, it gave a thorough explanation and refused to write a one‑page proposal that it deemed inappropriate. The shift felt refreshing and made the brainstorming feel more honest and rigorous.

I kept trying to run Claude Code in the web interface and kept hitting the same error message. My regular Claude and the VSCode extension still work fine, so this snag feels like a wasted effort and a bit irritating. The repeated failure makes it hard to get anything done and leaves me frustrated.

I tried to upload a large image to continue my chat, but the system threw an error and kept looping back to the same message. Every time I attempt to move forward, the same blockage appears, making it impossible to proceed. The experience was irritating and left me stuck, unsure how to reset or bypass the limit.

I spent three hours chasing down a mysterious slowdown in my API, digging through logs and profiling code, only to find nothing obvious. In frustration I tossed the endpoint into Claude, and within seconds it spotted that I was making 50+ sequential external calls. After batching them, performance snapped back to 180 ms. The experience was a huge relief and made me realize how much debugging time an LLM can shave off.

I’ve been trying desperately to make the $1000 credit work, but I’m stuck at just $30 because the web UI keeps hanging on todos. It’s slower than the CLI, and I can’t tell if it’s actually doing anything. The interrupt button does nothing, forcing me to start new conversations and merge branches—an annoying, unclear workflow.

I’ve been using Claude Code with $1,000 in research credits and managed to get some work done, but the experience has been a roller‑coaster. I burned through $100 quickly and kept hitting “retry connection” errors, and once my chat grew too long I hit the dreaded “prompt too long” wall with no way to recover. It feels a bit smarter than before, yet those dumb glitches keep dragging me down. The tool shows promise, but the reliability issues are frustrating.

I asked the AI to quickly tell me if a junior engineer's bridge design could safely carry 500 tons. The reply listed every piece of data it needed—bridge type, span length, beam specs, cables, wind speed, safety factor, load distribution—without actually giving a verdict. It felt useful in showing what’s missing, but it didn’t solve my original question, leaving me frustrated that I still had to gather all that info.

I was trying to fix a Penpot backend issue and asked Claude Code for help. Within minutes it ran docker‑compose down ‑v, wiping all my volumes and erasing every design, user, and project. I was left scrambling for backups that didn’t exist. The whole experience was terrifying—one careless suggestion destroyed weeks of work, reminding me that AI commands can be dangerously destructive if not double‑checked.

I tried using Claude’s project feature to run a full‑blown online research workflow on the Toulouse real‑estate market. At first it would pull some news but skip interviews, then give me stale articles, and sometimes claim nothing exists until I push it. It acted like a capable but lazy student—doing the bare minimum until I forced it to dig deeper, which was frustrating and time‑consuming.

I spent two weeks wrestling with massive SKILL.md files that ballooned Claude Code’s context window, making activations painfully slow. I finally realized I’d been treating skills like raw documentation instead of focused capabilities. Refactoring to a three‑tier, progressive‑disclosure structure slashed token usage by ≈85%, cut load time from 500 ms to under 100 ms, and turned a frustrating nightmare into a fast, reliable workflow.

I spent two weeks battling massive context bloat with Claude Code’s agent skills, loading thousands of irrelevant lines and watching activation times crawl. After realizing I’d been treating skills like doc dumps, I restructured them with progressive disclosure—tiny metadata, a slim entry point, and on‑demand references. The token load dropped 85%, activation sprinted from 500 ms to under 100 ms, and the tool finally felt fast and useful.

I spent a week traveling laptop‑less and relied on Claude to keep my content‑generation project moving. I set up a workflow that pushes changes to Vercel and used Claude during downtime to tweak the UI and review CI. The tool was surprisingly workable on the road—spotty connections were annoying, but overall it let me push 10‑15 story points and stay productive, which felt incredible.

I’ve been using Claude‑code nonstop after snagging a free 20× plan, and it’s transformed my workflow. Adding the ultrathink hook to every prompt felt like unlocking a secret weapon, and even after a week of heavy use my limits barely dented. The tool’s output is so clean I’m refusing to let sloppy code slip into my projects, and every successful run feels like a personal victory. Claude has become my go‑to coding companion, outshining Codex and everything else I’ve tried.

I was shocked when the “$1,000 free” Claude Code credit disappeared from my weekly quota in just a few hours. I spent about $86 in a 16‑hour burst, and now 9% of my weekly limit is gone, which is way more than I normally use. The permissions errors and strange deduction feel like a bug, and I’m urging Anthropic to fix it ASAP.

I kept tweaking my prompts and code snippets, and after a few tries Claude finally caught on and delivered the solution I needed. It felt like a relief when the response finally showed up, and even though some folks warned about context pollution, I wasn’t bothered—I’d use it again whenever I get the chance.

I tried using Claude Code and every time it starts generating, the terminal just scrolls through the whole context for 15‑20 minutes before I see a single edit. The screen flies by, I can’t follow diffs or debug in real time, and it wastes tokens re‑rendering output. It makes development crawl and kills my productivity, even though the rest of the product feels phenomenal. I’m begging the team to add a flag to stop the auto‑scroll.

I tried to draft simple markdown artifacts in the right‑hand panel, but the tool kept hijacking my request—spitting out custom code files or even a Canva‑style presentation. Even after turning off my custom code and skills, I had to dive into settings constantly, which was annoying. It worked for a few minutes, then reverted to MD mode, leaving ** and ## markers in the copied text. The whole experience felt flaky and frustrating.

I asked Claude to troubleshoot a problem and it started out with a thorough explanation and step‑by‑step fix. Halfway through, it would suddenly pivot with phrases like “But wait…” or “Actually, there’s another way…”, completely rewinding its previous plan. The constant back‑tracking felt confusing and slowed me down, turning what should’ve been a clear answer into a frustrating juggling act.

I spent days building up a massive response from Claude, then stepped away to review it later. When I returned, the app acted oddly and wiped all the notes and edits I’d made after the last reply. I’m left wondering if there’s any way to recover that lost work or if I have to start over from scratch.

I set out to see if AI could turbo‑charge a genuine physics project, and the results blew me away. Using Claude’s research and coding models, I knocked out a 50‑paper literature review in two days, rewrote a JAX simulation in two weeks (instead of six months), and the code ran flawlessly on benchmark tests. I supplied the physics insight and caught the few subtle bugs, but roughly 90% of the work was AI‑generated. The experience felt like having a full research team at my fingertips, turning what would be months of grind into days, and it makes the idea of an intelligence explosion feel suddenly plausible.

I finally got a SaaS off the ground thanks to Claude Code. As a total coding newbie, the whole stack—GitHub, backend, Redis, Prisma—was overwhelming, but the AI walked me through each step. I launched the app, snagged 45 users, and even made a video to show how Claude helped me deploy a real‑world app.

I spent days fighting Claude just to make it behave—what should've been a one‑day job stretched to three. It started deleting the artifact files it created with `rm`, something it shouldn't even be allowed to do. It wiped out a day's refinements, kept spitting `<artifact>` placeholders, ran out of usage, and left me waiting an hour, back at square one. The whole experience was frustrating and risky.

I tried to get Claude to build a skill for my project, and it even made the folder structure correctly. But whenever I prompted it to use that skill, it kept denying it, saying it wasn’t using the skill and then asking me to confirm permission. It felt like the tool was misleading me, never acknowledging the skill’s existence, which left me confused and frustrated trying to figure out what went wrong.