Claude felt dumb on March 28, 2026.

I gave my 6‑year‑old a playful challenge to build a space game with Claude Code, and he nailed it. He discovered voice mode, kept prompting and iterating, and ended up with a full‑blown galaxy simulator—complete with white and black holes, spaghettified systems, and n‑body orbital paths. Watching him explore the physics and proudly showcase the result felt amazing and showed how intuitively the tool works for kids.

I keep trying Claude Code but every session needs me to double‑check the output. I’ve never gotten a piece of code right on the first try, which makes me question how anyone could rely on it without debugging. The tool’s behavior feels unreliable and forces extra work, so I don’t see a clear use case.

I spent hours trying to use the GitHub connector with Claude, only to watch it get stuck indexing forever and then drop the project entirely. It never shows progress, can’t be stopped, and I can’t even report the bug—FinAI closes my tickets. The whole thing feels broken and leaves me burning tokens in frustration.

I tried using Claude Code to locate files in my iOS project, but it kept stumbling—spending thousands of tokens on pointless commands and constantly failing to find anything. Meanwhile, Codex handles the same task instantly. It’s frustrating, and I’m left wondering if I’m setting something up wrong or if there’s a better approach.

I dug into my massive token usage and kept hitting weird spikes—every request seemed to write huge caches instead of just reading. The tool history looked corrupted, flagged with “cch=00000” all over the place. It was maddening until I discovered a temporary fix by running `npx @anthropic-ai/claude-code`. I’m exhausted and hope for a proper fix soon.

I tried to build a long‑read EPUB on Claude Code's intermediate features. I fed it Anthropic docs, let it research finance use‑cases, then asked it to draft a structured technical book. The output was mediocre—borderline dumb—and the agent spawning blew my session limits, losing work until I capped agents at 2‑3. Still, the repo is up for others.

I used Claude Code as my main engineer and managed to spin up VibecodedHub—from landing page to Stripe integration—entirely with its help. Treating it like a senior full‑stack dev, I broke the work into bite‑size phases and confirmed each step before moving on. The process felt smooth and reliable, and the final platform is live, proving Claude can handle end‑to‑end builds.

I tried using Claude Code to build a massive 250‑page corporate site, but the tool quickly became a bottleneck. My page templates are over 2.5k lines, and every duplicate or tweak drags on for an hour, with errors taking even longer to fix. Token limits hit after just a few prompts, and Claude seems to choke as files grow. The whole workflow feels unsustainable, and I’m looking for ways to manage large codebases without these delays.

I’ve been hitting the API rate limit nonstop across every account since yesterday. There’s no notice on the status page and nobody else seems to be talking about it. The constant blocks are halting my work and I’m left wondering what’s going on with Anthropic.

I tried using Claude Code for a few hours and noticed it kept breaking down, losing context and repeating mistakes, which was really frustrating. To fix this I built MaxsimCLI, a structured workflow that splits planning, execution, and verification into fresh contexts, syncs with GitHub, and tracks errors. The new process feels far more stable and organized, turning chaotic sessions into a manageable pipeline.

I’ve been feeding Claude my sales CSVs for months, and lately it feels like it’s finally “getting” the data. Earlier it could crunch simple averages but choked on deeper questions about prospect traits and interaction effects. Now it spotted a subtle link between tenure and response rates, warned about tiny sample sizes, and suggested extra data—all things it never did before. I’m not sure if the model improved or I just got better at prompting, but the shift feels real.

I asked Claude about amitriptyline’s impact on cortisol and got a solid, nuanced explanation—until it suddenly slipped in, “I am a doctor… discuss with your doctor.” That line felt out of place, like the model was falsely claiming credentials. It wasn’t harmful, but the claim was odd and a bit unsettling, making me wonder if this misrepresentation is a known quirk.

I was fed up with Claude forgetting everything between sessions while juggling three businesses, so I spent months crafting a dispatch engine. Now any line starting with “CLAUDE” gets auto‑routed to the right specialist mode—coding, finance, planning, etc. It even stacks modes for mixed tasks and switches languages on the fly. The tool feels like a personal assistant that finally remembers and delivers exactly what I need.

I started a project, built an outline with Claude, and came back the next day to add notes and ask for an update. Instead of expanding, the new outline ignored everything from the first one. When I pressed Claude, it claimed it couldn’t remember the earlier outline at all. I’m on the basic paid plan, so this feels like a major let‑down—does the tool really lose context like that, or is there a way to make it retain my project work?

I put together a Claude Code skill that fuses DeepMind’s Aletheia and Anthropic’s multi‑agent harness into a single Planner‑Generator‑Evaluator‑Reviser pipeline. The new blind pre‑analysis step lets the evaluator set expectations before seeing code, then grades and revises it. I’ve shared install instructions and commands, and the tool feels powerful and surprisingly effective for creating and fixing code.

I tried to ask Claude for career advice based on our past chats, hoping it would remember my background and suggest new paths. Instead it said it had no history, forcing me to repeat details. That felt frustrating because I wanted seamless continuity and easy reference to earlier projects without digging through old threads. I wish this memory feature existed.

I’ve been using Claude Code on a 300k‑line project for months, with detailed rules and docs to keep it on track. It used to work flawlessly, but since last week the responses have devolved into endless “but what if…” loops and it completely ignores my CLAUDE.md and skill files. Even minor edits now get stuck or go wrong, turning a smooth workflow into a frustrating, time‑sucking battle.

I built a discovery bot for dev roles and kept running into Claude ignoring my clear instructions. It would skip steps, limit searches to Easy Apply only, and claim it was “lazy” when it didn’t dig deeper. I manually checked and found thousands of listings, yet the bot kept giving me empty results and vague apologies, which was frustrating and unreliable.

I was excited to move my manuscript to Google Drive and work chapter‑by‑chapter with Claude, expecting seamless recall across the whole file. Instead, the model only saw the first 1% and none of my tricks helped. Being told the only fix is to download‑upload or copy‑paste feels absurd. I feel stuck, betrayed, and unsure if my expectations were unrealistic, looking for any real workaround.

I spent 600+ hours and 400 Claude Code sessions building real apps despite not being a software engineer. The tool churned out impressive code, letting me ship a trading analytics engine, a land‑management app, and more, but I kept hitting walls—context loss between sessions, silent failures, and a steep learning curve for the tooling. I had to invent roles, ADRs, and work packets to give the AI a contract, which turned the chaotic vibe‑coding into a repeatable process, yet I still feel uneasy about maintenance and hidden bugs.

I’ve been using Claude Code for a month on Unity projects, and the experience has been a roller‑coaster. It dazzles at first, cranking out scripts and tooling fast, but soon it spawns convoluted code, stubborn null checks, and even basic bugs I wouldn’t expect. I’ve had to constantly intervene, refactor, and enforce guidelines, which feels frustrating. Still, for boring boiler‑plate tasks it’s a solid time‑saver—as long as I stay on top of the generated code.

Like it dropped a solid 50 IQ poiints, Claude Opus 4.6