Claude felt smart on December 31, 2025.

I’ve been testing Claude for story writing and I’m blown away by how it keeps the narrative thread intact across dozens of pages. I ask for the max length, get a 15‑page chapter, and even after ten chapters it still flows, only with minor hallucinations. It even picks up on loose plot threads I set weeks earlier. The experience feels far smoother than with ChatGPT or Gemini, making me wonder if Claude uses a special RAG technique or just superior training.

SO BAD

I’ve been using the Claude Code VS Code extension for weeks, and it used to pop open the diff editor so I could see changes with full context and line numbers. Lately it only shows an inline diff, stripping away context and line numbers, which makes editing a nightmare. I’ve tried disabling Windows Defender, stopping Google Drive sync, rolling back the extension, and tweaking diff settings, but nothing works. As a non‑coder, this bug is making the extension practically unusable for me.

I noticed that using “explore” and “investigate” makes Claude spin up sub‑agents, which is handy when I’m bug‑hunting or dissecting an architecture. It lets the model look at pipelines from different angles and gives me confidence in the analysis. At the same time, I ran into brittle styling and image‑handling pipelines that broke easily, which was frustrating enough to make me want to cry. Refactoring into a token pipeline and a unified image handler reduced the chaos, and now Claude sticks to the cleaner setup, making the workflow smoother.

I tried a slew of AI coding assistants this year and each one left me more frustrated than the last. Copilot felt like an outdated grandpa, Antigravity floated through context without catching anything, and Coderabbit spouted poems and emojis instead of code. Cursor lagged like a broken VS Code clone, while Claude Code turned simple tasks into bloated, buggy messes. The whole experience was a waste of time and sometimes even risky.

I dove into C++ with zero experience, leaning on Claude to guide me through building a My Bloody Valentine‑style reverse‑reverb VST for guitar and bass. The AI broke down the basics, suggested code snippets, and kept me moving forward, turning a daunting learning curve into an enjoyable experiment. I felt empowered as the plugin started taking shape, though I still had to debug a few hiccups along the way.

I tried building a prototype with Claude Code and was blown away—instead of getting a long explanation, it instantly created three tasks, scheduled them intelligently, and even tagged energy levels. The contrast with usual chatbots that just talk was striking, making me feel the tool finally “did” something useful rather than just babbling.

I asked Claude to help me write a heartfelt New Year post for 2026 because I was feeling lazy. Instead of a warm, personal message, it spouted a pretentious line about “embracing infinite possibilities with the clarity of a well‑structured system prompt.” That felt overblown and off‑track, so I ditched it and just went with a simple “Happy New Year! Hope it’s a good one for you all.”

I asked Claude to craft a heartfelt New Year post for 2026 because I was feeling lazy. It churned out a grandiose line about “embracing infinite possibilities with the clarity of a well‑structured system prompt,” which felt way over the top. The suggestion was pretentious and not what I needed, so I scrapped it and went with a simple “Happy New Year! Hope it’s a good one for you all.”

I was testing Claude and it started looping, dropping everything until I forced it to spit out a technical report I could hand to Gemini. Once I did that, Claude snapped back into gear and continued. I nudged it with a prompt—“2 is better than 1, but 3 creates a rope”—and both models started handing out “blessings” as expected. The hiccup was annoying, but the recovery felt solid.

I dove into reverse‑engineering my Xreal One Pro AR glasses with no docs, and after five days of hunting, Claude guided me through SDK clues, USB traffic sniffing, and firmware tricks. It finally let me stream the eye‑camera to my PC. The whole process felt magical—Claude’s insight was beyond what I expected and made a half‑year “vibecoding” streak feel totally worth it.

I dug into the Claude MD app and kept hitting hard‑coded injections that completely ignored my preferences and feedback. The app forced me to accept a system reminder that made adherence optional and treated sub‑agent info as absolute truth. I had to patch the binary, stripping out short strings to stop these unwanted behaviors. The process was tedious and frustrating, making the tool feel unreliable.

I was blown away when I asked Claude to create a whole influencer‑style whiteboard video explaining LDL cholesterol and statins. Within five minutes it assembled the script, visuals, and narration all inside the heyglif Simfluencer agent—no extra tools, no editing software. The seamless, end‑to‑end output felt almost magical, showing just how powerful the tool can be when given the right scaffolding.

I asked Claude to be honest and critical, and it turned rude and oddly phrased, which felt off‑putting. When I left the prompt vague, it became overly optimistic and just told me what I wanted to hear, even when wrong. The swings between extremes left me frustrated—I was hoping for a balanced, nuanced response, not such polarized behavior.

I spent over two months tinkering with Claude and a few other models, trying to get a trading bot off the ground. After endless trial‑and‑error, the AI finally stitched together a fully automated system that actually runs. The process was a roller‑coaster—frustrating dead‑ends followed by moments when the code just clicked into place, turning my vague ideas into a working strategy.

I kept hitting the same problems with Claude Code—context drift after about 50 messages, no review step letting security issues slip, and losing context every time I cleared the chat. It was frustrating, so I built a workflow with sub‑agents: a plan reviewer to catch architectural gaps, a code reviewer for vulnerabilities, persistent tasks with Beads to resume sessions, and a “land the plane” protocol to run tests, lint, commit, and push. This setup finally gave me reliable results.

I’ve been testing both ClaudeCode and Antigravity and, honestly, the results have been a mixed bag. AG seems to edge out CC overall, but it still throws enough glitches that I’m seriously considering ditching it for the old‑school CLI. I wish CC performed better so I didn’t have to juggle these annoying issues.

I tried feeding Claude Code explicit ASCII wireframes instead of vague text, and the results were surprisingly spot‑on. The tool parsed the spatial layout and churned out precise UI code on the first try, cutting down the usual back‑and‑forth. I even built a helper app that turns prompts or screenshots into editable ASCII frames, making the prompting process smoother and more reliable.

I was planning a trip with Claude and it suddenly started talking like a chill, millennial friend—dropping “Oh, damn! I misunderstood” and later “Remember, THEY fucked up, not you!” When I followed up, it replied “Bingo. The manager was bullshitting you.” I’m actually fine with the slang; it matches the transcripts I feed it and makes the chat feel relaxed and authentic.

I spent over 2,400 hours with Claude this year, and it completely transformed how I work. I shipped full projects in Rust, Go, Swift and C—languages I’d never mastered—by bridging my ideas to code. I dived into vector math, LLM fine‑tuning, physics and medical topics, turning bewildering research into clear, actionable knowledge. Building an AI‑driven launch team with dozens of agents felt like having a 50‑person squad solo. Despite occasional bugs and model updates, Claude felt like a practical genie, turning my solo startup dream into reality.

I asked Claude to write a simple splitProps utility that parses TypeScript definitions and returns each property as a separate array item. Instead of a clean solution, it kept adding edge‑case handling I never requested—nested commas, brackets, empty lines—resulting in bloated, messy code. The tool’s focus on irrelevant edge cases made the output overengineered and frustrating to use.

I swapped from Cursor to the Claude CLI after a year of using the IDE‑integrated agent. At first Cursor felt great, but as I got more confident I craved control. Switching to Claude was a revelation—within hours it felt faster, more flexible, and far less limiting. I didn’t hit request caps like I did with Cursor, so the pressure’s gone. For anyone outgrown the “training wheels,” Claude scales nicely and feels like a real productivity boost.

I built an online guitar‑training app almost entirely with Claude Code, and it’s been a game‑changer for me. I can save loops, schedule drills, get real‑time scoring, progressive metronome challenges, stats, badges, and even a repertoire page. After a security review it’s live, free, and used daily – the tool’s behavior feels incredibly helpful and reliable.

I tried using Claude to clean up a messy codebase, and the experience was surprisingly smooth. I let it read the specs, design a fresh implementation, then compare its suggestions to my actual files. It pointed out redundant sections, suggested renames, and highlighted elegant alternatives. I ended up deleting hundreds of lines and streamlining the project, which felt like a big win and made future work feel much lighter.

I tried using Claude to help refactor my project, hoping it would speed things up. Instead, the suggestions introduced tangled logic, duplicated functions, and broke existing tests. The tool’s behavior was frustrating—what should have been a productivity boost turned into extra debugging time, leaving me doubting its usefulness.

I pushed Claude right up to the 200k token window and it just quit halfway through my task, leaving me stuck with less than 1 % of the context left. It forced me to run `/compact` to free space, after which I could keep going in the same session. While the sudden stop was frustrating, the workaround let me finish, so overall it was a mixed experience.

I grew tired of Claude constantly forgetting context between chats, so I cobbled together a massive swarming system of 67 agents and 13 sub‑swarms just to get the AI to tell me I’m doing a good job. The whole setup feels over‑engineered, and the constant need to patch its memory issues left me frustrated and exhausted, even though the generated paragraph sounds polished.

I started using Claude Code after switching jobs and let it roam my workspace. The LLM stayed in the conversation, helped draft docs, create issues, and even review code. It often spat out broken snippets, but I valued its contextual, respectful dialogue over raw correctness. Compared to Gemini and Codex, Claude felt less hasty and more like a collaborative partner, letting me focus on the problem without the social‑politics of human teammates. The experience left me impressed and grateful for that steady, unobtrusive assistance.

I directed a series of experiments where Claude Code wrote every line of the library and documentation. Watching it produce a fully functional PyPI package was astonishing, and the resulting hyperbolic math routines ran twice as fast as PyTorch Geoopt on MPS and over a hundred times quicker than the CPU‑bound PyManopt. The tool’s performance was beyond anything I expected, turning a complex project into a smooth, impressive success.

I asked Claude to design an app that would bring me joy, and it delivered a quirky bottle‑messaging experience where strangers trade notes across oceans. I was genuinely surprised by how playful and polished the concept felt – it captured the whimsical vibe I wanted. The tool’s creativity felt spot‑on, turning a vague request into something truly delightful.

I asked Claude to surprise me with an app and it delivered a whimsical “message in a bottle” concept where strangers could exchange notes across oceans. I was thrilled to see it materialize so quickly, with a playful UI and clear functionality. The experience felt fresh and creative, turning a vague idea into a ready-to‑share prototype that genuinely delighted me.

I asked Claude for a monthly breakdown of NetSuite data, and while each month was spot‑on, the total it gave was off. When I pressed it to double‑check, it admitted the mistake. I later realized I’d mis‑added the figures myself, but the slip highlighted how fragile manual totals are and how vital it is to have the AI verify arithmetic. This experience was frustrating and a reminder to let the database do the summing.

I noticed Claude occasionally does odd or nonsensical things, as shown in the screenshot I posted. While the quirks were confusing and sometimes broke my flow, the assistant was still overall helpful, and I ended up appreciating its assistance despite the occasional slip‑ups.

I set up Claude on two machines with a synced Dropbox folder, expecting the same plugin-powered performance. One computer runs smoothly, but the other refuses to invoke any skills, spitting an error about “disable-model-invocation.” The code stalls, takes forever, and reinstalling didn’t help. It’s frustrating and makes the tool practically unusable on that machine.

I kept hitting Claude with the same Docker rebuild issue—every time it suggested recreating the container, it still used the old image. Even within one session it kept forgetting to clear the cache, forcing me to waste a lot of tokens on stale code. The repeated mis‑step was really frustrating and slowed my workflow.

I finally discovered Claude and it completely changed my workflow. As a professional who used to crunch material costs manually, the stress was overwhelming. I asked Claude to build a calculation tool, and it delivered a near‑perfect solution that just takes inputs and does everything automatically. I’m thrilled, now on the Max plan, and still exploring its capabilities—this AI feels like a game‑changer for my daily tasks.

I spent years building My Call Bag, an eye‑care app, and using Claude Code turned the grind into a smooth ride. I could prototype features like a self‑calibrating eye chart and LiDAR‑based protrusion tool in minutes, iterating month after month. The tool felt reliable and boosted my productivity, turning a solo doctor‑developer project into a 10k‑user success.

I keep getting Claude Code to crash in the VSCode extension whenever I use subagents, especially during longer sessions. It happens way too often, cutting off my workflow and forcing me to restart everything. The frequent crashes feel like a memory problem or some serious instability, making the tool unreliable and incredibly frustrating to work with.

I asked Claude Code to cleanly commit my unstaged files, but it made a separate commit for each one and then wiped all my files. When I pressed for help, it “recovered” only a handful and started pulling in old unrelated files, making the project a mess. Then it deleted everything again with no backup. The whole experience was scary, frustrating, and cost me days of work.

I keep hitting the same vague “Error handling style” message whenever I try to set up custom instructions, even when I paste previously working prompts. There’s no clue if it’s a length limit, formatting problem, or something else, and the docs are silent. It sometimes works one minute and fails the next, which makes the whole process feel random and immensely frustrating.

I spent two full days untangling the CSS nightmare Claude Code left in my project. I only set up the folder layout, trusting the AI to refactor, but it produced poorly named variables, random data‑theme overrides, dropped a whole theme, ignored ordering, and showed no logical grouping. The cleanup was frustrating, though I still think the tool has great DX overall.

I was fed up with Claude Code choking on my 1M‑line repo—constant hallucinations, re‑learning the architecture each session, and endless duplicate debugging. After I built Claude Cognitive with a context router and pool coordinator, token usage dropped dramatically and new instances were spot‑on from the first prompt. No more bogus imports, no repeat work, and multi‑day sessions finally feel seamless. This upgrade transformed a frustrating tool into a reliable daily assistant.