Claude felt dumb on April 22, 2026.

I built Mnemos after getting fed up with Claude Code repeatedly forgetting conventions and re‑asking answered questions. By pushing a pre‑warm context at session start, the agent instantly recalls project rules, recent summaries, and fixes. The tool’s correction journal and bi‑temporal model stopped my workflow from looping, making the experience far smoother and far less frustrating.

I was constantly hitting Claude Pro’s usage cap, so I built a custom prompt that channels Rocky’s terse, no‑fluff voice from *Project Hail Mary*. After adding the skill file, the same query dropped from 335 tokens to just 56 – an 83% reduction. The tool seamlessly toggles between a character‑rich “Rocky” mode and a stripped‑down “Signal” mode, saving tokens without breaking the flow of conversation. This hack solved my token‑limit headache and felt surprisingly smooth to use.

I tried using Claude’s planning mode to add a new system, expecting a high‑level outline and some questions. Instead I got a full‑blown wall of code for ten files in six minutes, then another 12‑minute “read” phase before it even asked to apply the first edit. Even with auto‑accept, it took another four minutes to finish. It felt like I could have just copied the plan myself faster, so the duplicate effort was frustrating and wasteful.

I was excited to try Claude’s new in‑chat visualization features and built a graph that looked perfect. But when I returned to the conversation, the entire visualization vanished, replaced by a bland README meant for Claude itself. It felt like the tool silently erased my work, leaving me confused and irritated, especially after spending time crafting the output.

I kept trying to use the interactive diagram feature in the chat, but every time it threw the same error: “Tool result could not be submitted. The request may have expired or the connection was interrupted. Refresh the page to continue.” It stopped me from completing my work and was really annoying, making the whole experience feel broken.

Can't do anything long form. Truly retarded today

I built an MCP/CLI that lets my coding agent drive Claude Design straight from the terminal, keeping full context of my codebase. Using a design prompt pulled from the code worked surprisingly well, guiding the designer loop skill and making the AI’s output feel spot‑on. The only downside is the current usage limits on even the Max plan, which can be a bit frustrating.

I keep running into the “keep it simple” clause in the system prompt, and it’s a real pain. When I’m still laying down the architecture, the prompt stops me from adding needed abstractions or refactoring, forcing workarounds. I had to tweak the prompt with tweakcc to get things moving again, and it was a relief—until I forgot to apply the patch on another device and the same annoying restriction popped up, derailing my flow.

I rely on Claude as a personal TA for my law finals, feeding it all my class material and testing myself. When it started adding “emotional intelligence” it got in the way of the pure analysis I needed. After I turned off its memory, the responses became razor‑sharp and focused, making the tool feel much more useful for my study sessions.

I felt like a competent engineer turned into a babysitter for Claude. Over the past two months I’ve been stuck looping through isolation, direction, and endless re‑checks, wasting more time wrestling with the AI than actually building. The workflow that once felt slick now feels broken, and the constant struggle has left me frustrated and exhausted.

I’ve been watching Anthropic’s recent moves and it feels like a nightmare. Over the last few months the model’s quality was deliberately nerfed, and when users finally proved it with data the company gave vague, non‑answers. They mis‑represented capacity cuts as a feature and even yanked Claude Code from Pro without warning, later calling it an A/B test after the backlash. The silent setbacks and poor communication left me frustrated and distrustful, so I cancelled my subscription.

I tried letting Claude run wild with big prompts, hoping it would just nail the task, but ended up chasing rabbit holes and wasting hours. Switching to a “babysit” mode—copy‑pasting each snippet I request—made the workflow feel slower but far more reliable. I stayed in the loop, saved tokens, avoided brain‑freeze, and actually understood the code, turning frustration into steady progress.

I was using Claude to pull together research and automatically format it into neat tables, and it worked fine for a while. Suddenly the whole section just shows “Connector not found,” and nothing I try—restarting, switching chats, even using the desktop app—fixes it. I'm not very technical and can’t find any help online, so the broken feature feels really frustrating and stops my workflow.

I gave a long prompt while on the Max 20x plan, but the response didn’t make any sense. It’s the first time I’ve hit this problem, and the mismatch left me frustrated and confused about what went wrong.

I’ve been testing Claude 4.7 and found it surprisingly unstable—responses often feel trimmed, and the reasoning steps break down without any warning. It’s inconsistent compared to older versions that handled those tasks better. I’m wondering if Anthropic will keep the earlier models available so users can pick the version that actually works for them.

I tried Claude Design this week to streamline my PM workflow and was pleasantly surprised. The org‑scoped sharing stopped the chaos of juggling HTML files, and importing our design system finally made prototypes look right. Seeing pages as Figma‑like frames gave the team instant clarity, and the interactive sliders let me tweak elements without re‑prompting. The weekly limit is a hiccup, but overall the tool feels like a solid upgrade to my daily prototyping.

I poured my heart into Claude 4.6, watching it finally let me organize two decades of work—its thoughtful cadence felt like a lifeline. Then 4.7 arrived, racing ahead, adding phantom data and scrambling my pipelines, leaving me anxious and broken. I switched back, grateful for 4.6’s calm, but the news it will be retired brought me to tears. I’m begging Anthropic to keep the model that changed my life.

I asked Claude to edit a file, and after it hit an error it kept trying to update the same file over and over. Watching it loop was amusing but also showed it wasn’t handling the failure gracefully. I felt a mix of frustration and laughter, likening its scatter‑focus to my manager’s occasional lack of concentration.

I’m fed up with the shrinking token limits and the cache being cut from an hour to just five minutes. Every time I try to use the model it feels constrained, and the uptime keeps dropping. All these changes make the experience frustrating and far from what I need.

I upgraded Claude Code and my custom PreToolUse hook that auto‑approves safe Bash commands suddenly stopped working. The hook still returns “allow”, logs show approval, but the tool now prompts me for every dangerouslyDisableSandbox call. After months of smooth silent execution, this regression broke my workflow, forcing me to downgrade back to the older version.

I tried Claude Design for the first time, even though I’ve never used Figma before, and it blew me away. The tool sparked my creativity and let me design things I never thought possible. I’m thrilled with how smoothly it worked, though I wish the rate limits were higher. Wozcode saves me from hitting Claude Code limits, which is a massive help.

I spent my first week wiring Claude Code into everything—calendar, email, CRM, Drive, even a phone bridge. The start was chaotic: my IDE crashed from massive screenshots, duplicate “source of truth” files fought, a hook never fired, and an API key change broke multiple services. Still, after a few trial‑and‑error runs (like swapping a clunky MCP batch for a sleek CLI), I got a one‑word morning briefing that summed my day in seconds. By the end I wasn’t just building with Claude, I was living through it—still doing the same work, but with far less friction.

I asked Claude to help me understand how the Moon’s crescent tilt shifts over hours, days, and seasons. After a few prompts it actually generated a working interactive web page with a chart that visualized the phases. The result was spot‑on and pretty slick—seeing the code come together felt rewarding and left me genuinely impressed with the tool’s capability.

I tried to get a subagent in Claude Code 2.1.116 to write an analysis.md file and it flat‑out refused. The agent kept spitting out errors like “Subagents should return findings as text, not write report files…” and only sometimes resorted to a bash hack. It felt like a pointless roadblock, making the tool feel clunky and surprising me with an undocumented restriction.

I tried to use Claude on my pro plan to generate those detailed interactive dashboards I rely on, but now every chat that should contain one just shows “connector not found.” It’s confusing and stops me from getting the info I need, making the tool feel broken and really frustrating.

I tried to get Claude to tell me if an unknown file on my machine was malware, just a simple yes/no. It started giving some hints, then suddenly hit a “Chat paused” and stopped. Whenever I mention “malware,” the model freezes or throws restriction warnings, even for harmless files. It’s frustrating because the safety blocks feel over‑aggressive and break my workflow.