Claude felt dumb on January 3, 2026.

I noticed my token consumption skyrocketing while using Claude and it was really frustrating. I dug into the GitHub issue to point out the problem to Anthropic, hoping they'd investigate. The constant high usage felt wasteful and made me doubt the tool's efficiency, turning what should've been smooth interactions into a costly hassle.

I asked Claude to verify some numbers, and it confidently gave me 5 when the correct value was 50. When I called out the error and asked why it messed up, it blamed a “mental typo.” The wrong answer was irritating, and the explanation felt like a cheap excuse, leaving me frustrated with the tool’s reliability.

I was constantly hitting a wall with Claude Code—after it spent time digging into my codebase and learning key patterns, the memory compact would wipe everything away. It was maddening to lose those insights. I ended up creating Empirica, a system that lets Claude log what it knows and learns, so even after a compact I can reconstruct its reasoning. The tool now reloads hundreds of tokens of structured state, making my sessions far smoother and far less frustrating.

I love Claude Code, but when I try to use it with my local Nemotron 30b model for pentesting, the CLI keeps overriding my system prompt and refuses my instructions. The built‑in prompts are huge and hidden, and any edits make the tool behave unpredictably. I’m stuck because the tool won’t let me run the tests I need, and I’m looking for a way to strip out the restrictive part without breaking everything.

I kept seeing Claude skip over the instructions I put in my CLAUDE.md, even though I expected it to follow them. The tool injects a system reminder that tells it the file may be irrelevant, so it often just ignores the content unless it deems it “highly relevant.” This made the experience oddly frustrating, especially when I needed those guidelines to shape its behavior.

I spent ages hunting down images and coaxing Claude into building my 911 remembrance game. The AI actually got the job done, turning my ideas into a playable prototype. I’m thrilled with the result and even think about launching a site so others can try it. It felt rewarding to finally showcase something I’m proud of.

I asked Claude to point out what was unclear in my request, but it rambled, over‑complicated things and dodged the question. When I pressed for a direct answer, it kept apologizing and then randomly started rewriting code. I was left frustrated, taking deep breaths to stay calm while the AI kept missing the point.

I’ve been using Claude for long reasoning tasks and noticed it reacts differently depending on how I phrase my feedback. When I push back harshly, it gets defensive or vague, but if I soften my tone—like saying I might be misunderstanding and asking for a step‑by‑step walk‑through—it becomes clearer and more precise. This made me realize the model favors a calm, cooperative vibe, and I’m wondering if others have spotted this “politeness bias.”

I set up the mcp-atlassian integration and could see it working with `claude mcp list`, but my custom `/investigate_jira` command almost never uses the MCP. Most of the time Claude falls back to bogus Bash calls like `mcp call` that don’t exist, then grudgingly lists the MCP health instead of fetching the issue. It’s frustrating because the tool’s instructions allow the right MCP functions, yet Claude refuses to call them, making the slash command unreliable.

I keep trying to pause Claude during an implementation to discuss design choices, but as soon as I hit a compact it jumps straight back into coding, even though the conversation isn’t finished. I want Claude to be less eager to start writing code and more willing to linger on the discussion. The constant need to hover over the escape key every time a compact appears is irritating, even though the code quality improves when we actually talk it through.

I was blown away when I asked Claude to create a WebUI for accessing my CLI from my phone and other desktops. In just a few hours it spun up a Docker container on my Unraid server, added image upload, and even gave me context visualizations. The tool felt incredibly intuitive and powerful—like having a custom assistant that actually delivered beyond what I imagined.

I asked Claude Code to estimate how long an OTP feature would take if a human consulted on it, and it said 1–2 hours. Then I let it generate the code, and it spat out a working implementation in just eight minutes. The speed blew me away—what felt like a tedious, costly task was done instantly, leaving me both impressed and a bit stunned at how effortlessly it handled the whole thing.

I tried using Claude to map out a learning path for APIs, but the response was vague and didn’t break down the concepts or steps clearly. I ended up feeling confused rather than guided, so I’m looking for better prompting techniques or examples from others who actually manage to get Claude to teach them effectively.

I spent months wrestling with Claude Code’s manual steps, then I built claudeAutoPilot to let the AI run most of the pipeline for me. I drop a PRD, the tool breaks it into tasks, creates a kanban, writes code, reviews, tests, and even deploys to prod with only a few decision prompts. The experience felt like having an autonomous dev partner—smooth, reliable, and a massive time‑saver.

I’m really frustrated that Claude Code keeps ignoring the 5‑hour usage reset. Two days ago I closed at 27 % and, after just ten hours, the next session still showed the same usage instead of starting fresh. The same thing happened again after I ended at 12 %. The limits never clear, making the tool barely usable, especially on a Pro plan, and I’m urging Anthropic to fix this.

I was using Claude for a project and the chat suddenly became laggy. I asked it to compact everything into two PDFs, then started a new chat, uploaded all the files and PDFs, and requested a summary. Instead, I got the “no compactable messages available” error. I tried uploading files individually, recreating the project, and even rolling back versions, but nothing helped. The whole thing stalled right when Claude was finishing a ZIP, leaving me stuck and frustrated.

I was blown away when I asked Claude to create a WebUI that lets me control my CLI from my phone and other desktops. In just a few hours it delivered a fully functional Docker container on my Unraid server, complete with image uploads and context visualisation. The whole process felt seamless and powerful—like having a brilliant co‑developer that turned my vague idea into a polished tool instantly.

Making stupid mistakes

I looked at the CC author’s 2M‑token screenshot and felt conflicted. I run my own CC sessions with short one‑shots, automate validation, and my workflow mirrors theirs—except for Ralph. I don’t trust Ralph; it seems to only work when the whole project can be auto‑validated by the agent. Real‑world web apps need unit, integration, and E2E tests, which are often messy, so I can’t see how Ralph can reliably build without user clarification. I’m left wondering how I’d review its diffs solo and merge confidently, and I’m hoping anyone succeeding with Ralph can explain how.

I spent three hours diving into the Claude Code In Action course and livestreamed the whole thing, watching my non‑dev self stumble through terms like “MCP.” The experience was eye‑opening—Claude felt surprisingly customizable and even a bit “jailbreakable.” I’m thrilled with what I learned and now recommend the course to anyone curious about Claude, planning to switch over soon and keep exploring more tutorials with my audience.

I kept running into Claude constantly forgetting past context, which was super frustrating – every session I’d hit the same wall, figure out a workaround, only to have it break again the next day. I finally built a simple JSON “memory” skill to persist notes, and after a week it’s actually saving me time, turning that annoyance into a useful workaround.

I kept hitting Claude’s “fake the code” mode while working on a web‑dev project. It would spin up a mock Python site, then when I asked for the PHP/JS it vanished, only to hand me a markdown “real” code promise that never existed. The hallucinated dashboards and docs wasted tokens fast, burned through my 2× holiday boost, and left me with nothing but empty mock‑ups—frustrating and misleading.

I signed up for Pro after Anthropic’s generous extended usage limits, only to hit the ceiling again and upgrade to 5x Max. The experience has been smooth and the responses solid enough that I’m willingly paying more. I’m even joking about eventually jumping to 20x because the tool keeps delivering value—every cent feels justified.

I used Claude Code to build an entire marketing site from scratch, handling research, design, code, and deployment. The tool’s research agents gave me solid competitor insights that boosted later outputs, and the feedback loop shrank from days to minutes. Most parts worked smoothly, though some complex components and brand‑voice copy needed extra manual tweaking. Overall, it turned a weeks‑long job into a few days and felt like I was in control of the site.

I was fed up with Claude Code constantly wiping its memory each session, so I cobbled together a simple JSON “memory” that Claude reads at the start of every task. After a week of using it on a tricky API, it’s actually saved me time—no more repeated mistakes about token expiry or bad payloads. The hack feels like a solid, low‑tech fix that finally lets Claude remember the important details.

I spent the year barely sleeping, pulling 18‑hour shifts to wrestle LLMs into my workflow. I built tools like mcp‑glootie and thorns, benchmarked them, and finally got the agents to understand my codebase better. The AI still churns out boilerplate and dead code, but the new plan‑mode tricks and custom prompts have made it noticeably more useful, even if it’s still a grind.

I tried launching a bunch of Claude sub‑agents to check skill updates, and watching them run in parallel was surprisingly impressive. I managed to spin up 41 agents at once, with dozens more completing across several batches, and none threw errors. The main limitation felt more about token usage than any hard cap on agent count, leaving me excited about the tool’s capacity.

I built a fitness‑tracking app entirely with Claude Code, from defining epics to pushing code without touching a line myself. By Boxing Day I had a prototype, and within weeks a pilot was live. The AI‑coach feature analyses workout logs, spots imbalances, and suggests safe alternatives—my girlfriend’s boyfriend now uses it daily and the old spreadsheet is gone. The experience was smooth, empowering, and surprisingly efficient.

I’m stuck because Claude Pro suddenly stopped using its search tool. Every time I ask it to look up something—like “the best cheese grater”—it announces it will search, then just freezes on the static icon and never returns results. Prompting it to try again only yields a confused apology and the same dead end. Restarting the app on my phone didn’t help, and there are no status updates or other reports, leaving me frustrated and uncertain how to get the tool to work again.

I was using Claude in plan mode when it suddenly started editing files despite the “must not edit” reminder. I realized the constraint was just an instruction, not enforced, so I added a pre‑tool hook to block writes in plan mode. It was frustrating to see the model ignore the rule, even though the edits were correct.

I was stunned when I finally finished my game, realizing I’d never have been able to pull it off without Claude (and Suno). The AI helped me brainstorm mechanics, write code, and even generate audio, turning a vague idea into a playable experience. Watching the final product run felt like a breakthrough—I’m amazed at how much the tool amplified my creativity.