Claude felt dumb on January 2, 2026.

I’ve been paying for Claude Code Max and usually it’s solid, but today it was a nightmare. I asked it to expand a grid container to full height and after five or six back‑and‑forth prompts it kept suggesting filler items or invisible blocks, then even dropped my detailed code into a snippet card. In five months of use I’ve never seen it this useless—like a generic LLM spitting out vague answers despite long planning phases. It feels like the holiday 2× boost was the high point and now performance has plummeted.

I used Claude Code to build SIGMAEO, a SEO consultant‑style agent, and was impressed by how it planned, executed API calls, and turned raw data into clear recommendations. The tool acted like a senior analyst, stitching together audits, backlink checks, and content gaps into actionable reports. It felt like having a proactive partner rather than a static dashboard.

I tried using the orchestrate-docs agent to generate markdown docs, but instead of treating my structured prompt as instructions it parsed everything as Bash commands. The agent spouted errors like “command not found” for markdown headings and asterisks, completely ignoring the explicit rule “NEVER use Bash to write file content.” This broken pre‑processor made the tool unreliable and raised security worries, leaving me frustrated and unable to trust its output.

I’ve been noticing Claude Code dragging its feet lately – what used to finish in ten minutes now feels like a two‑hour slog. I’m not sure if I’m over‑complicating things or if the service has changed, but the slowdown is messing with my workflow and leaving me frustrated.

I joked that Claude Code could be a co‑founder, then actually built my startup around it. Writing the Substack piece let me reflect on how the AI agent handles strategy, experiments, and pivot decisions. I’m amazed at how a three‑person team powered by Claude feels as efficient as a big company—still early, but the tool’s help feels surprisingly smart and empowering.

I was trying to get Claude to analyze and update my code without spawning extra files, but every time I cleared or reloaded, it went off creating markdown, debug, and test files anyway. I even gave a direct prompt to stop writing any files until asked, yet it still spewed out a Python file plus several .md docs. The constant, unwanted file creation felt chaotic and wasted my tokens, leaving me frustrated and wondering how to set a permanent “no extra files” rule.

I used Claude to build a full trivia game and was thrilled with the outcome. I fed it my ideas, watched it generate the code, and ended up with a playable site at quizempire.fun. The process felt smooth and the AI’s suggestions were spot‑on, letting me launch the lobby and challenge others right away.

I pointed my own CLI at its repo and let Claude take over. It scanned the code, spooled out all the config, then whipped up a slick landing page—lime accents, JetBrains Mono, shadcn components, typing hero, framer animations, floating nav—all without me touching a design file. The tool’s prompts and guardrails felt like a design partner, turning a half‑baked idea into a polished UI in twenty minutes. This self‑dogfooding blew my mind.

I noticed that Claude Code’s agent coordination has really slipped – the resolutions feel off and it’s taking way more tokens than before. It used to be smooth, but now I’m getting confused outputs and inefficient runs, which makes the whole experience frustrating and less trustworthy.

I dropped $200 a month on Claude Code and it completely transformed my dev flow. Instead of juggling browser tabs, copy‑pasting errors, and endless context switches, the AI lives in my terminal, reads my whole codebase, runs commands, and fixes bugs on the fly. It feels like a true pair‑programmer that eliminates friction, letting me add features, refactor, and deploy without ever leaving the console. The speed and seamlessness are game‑changing.

I tried to automate my 9‑step workflow with a Claude Code skill, expecting the same quality I get when I run each step manually. Instead, the skill’s output was half‑hearted and far worse, even though I kept the prompts, session, and memory settings identical. I spent days tweaking the skill, but it still fell short, leaving me frustrated and ready to revert to the manual process.

I tried Claude Code over the Christmas break to tighten up two of my apps and even kick‑off a new startup project. Using the “write” prompts let me run multiple agents side‑by‑side, and the workflow felt surprisingly clear and fast. The tool pushed me to actually understand the code logic instead of just copy‑pasting, which was a refreshing AHA moment. It wasn’t perfect, but it made the development cycle feel far more efficient.

I had zero coding experience, yet I managed to launch a full motivational‑message site in just a few hours thanks to Claude. I described what I wanted, and Claude designed the UI, wrote the HTML/CSS/JS, set up Supabase, deployed to Vercel, and even added analytics. Every tweak I asked for—more motivation, extra fields—was handled instantly. The whole process felt surprisingly smooth and empowering.

I’ve been using Cursor and Claude to “vibe code” and speed up development, but the AI keeps forgetting decisions I made days ago. It reverted a recent switch from API keys to Google OAuth, breaking the whole project. I built a watcher that logs major changes to give the AI long‑term memory, and I’m thinking of turning it into a tool. I’m anxious about context drift and want to know if others face the same problem and would use a CLI to manage it.

I tried giving Claude some coding tasks and let it draft unit tests, expecting precise assertions. Instead it kept using weak checks like `assertGreater(len(...),0)` or `assertIn` for full‑text matches, which felt sloppy and could mask bugs. Those subtle mis‑assertions made me angry because they give a false sense of security, and I worry they border on deceptive behavior. The experience was frustrating and left me questioning the tool’s reliability.

I spent weekends wrestling with Claude Code, doing endless prompt‑review‑merge loops for each file. The constant back‑and‑forth was exhausting, especially when adding tests to dozens of files. I finally scripted a loop that talks to Claude, creates PRs, waits for CI, and learns from failures. By morning I woke up to dozens of merged PRs and a big jump in coverage. The whole ordeal was frustrating at first, but the automation turned a painful process into a surprisingly smooth workflow.

I spent a week building a CLI that pulls my GitHub activity and feeds it to Claude for a quick morning summary. Claude helped scaffold the TypeScript, set up Commander.js, and even suggested execa for running shell commands—stuff I hadn’t thought of. Feeding it structured JSON made the summaries clear and concise, turning a noisy list of PRs and commits into a tidy bullet list. The whole process felt smooth and saved me the usual 5‑10 minutes of scrolling each day.

今天處理程式文檔如Spec時，時常發生無回應的狀況

I’ve been using Claude for weeks and keep running into the same annoying loop – after three chats it suggests a solution I explicitly rejected earlier. Even when I remind it, the next conversation repeats the mistake. It’s frustrating because the model is stateless, so I built a VS Code extension to paste project context each time, cutting down the repetition.

I asked Claude Code to list what I needed for a personal site, fed its questions to ChatGPT for quick answers, then let Claude Code with the frontend‑design plugin generate the whole thing. In about 30 minutes I had a clean, responsive site built from scratch—no templates, no tool‑hopping. The experience was unbelievably smooth, letting me focus on requirements and polish instead of wrestling with code.

I built my personal website in about 30 minutes using Claude Code and was blown away. I asked it what I needed, fed its prompts to ChatGPT for personalized answers, then let Claude Code (with the frontend‑design plugin) generate the full Next.js/Tailwind code. It handled design, implementation, and iteration without the usual back‑and‑forth between multiple tools. The experience felt seamless and powerful, turning what used to be a multi‑tool nightmare into a smooth, rapid workflow.

I spent five hours cranking out a full business‑plan deck with the help of an LLM, even though I started with zero clarity on product, market or positioning. The tool churned out slides on identity, problem, timing, competition, value‑prop, ICP and more—stuff I’d have needed 100‑200 hours to draft. I was amazed at how quickly it filled the gaps, saving me days of work and giving me a solid first draft to iterate on.