Users circulate Claude workarounds after reported coding and reasoning nerfs

Claude users spent weeks feeling like the model had gotten dumber at coding. Not just a little worse — worse in the specific, maddening ways power users notice first. More repetition. More lost context. Less willingness to think through a plan. The news is that Anthropic now says much of that wasn’t a hidden model downgrade at all. It was a stack of product changes and bugs in Claude Code and related tools that landed close together. ### What actually broke? Anthropic’s April 23 postmortem points to three separate changes. On March 4, Claude Code’s default reasoning effort was cut from high to medium to reduce latency. On March 26, a session-memory change meant to trim old thinking after an hour idle kept firing every turn in some sessions. Then on April 16, a prompt tweak to make Claude less verbose ended up hurting coding quality and was reverted on April 20. (anthropic.com) ### Why did users read that as a nerf? Because from the outside, that is exactly what it looked like. If a coding model starts answering faster but with shallower plans, forgets what it was doing, and gets more repetitive, users do not experience that as “three interacting product regressions.” They experience it as “the model got nerfed.” Anthropic says the API and inference layer were unaffected, but Claude Code, the Claude Agent SDK, and Claude Cowork were affected on different schedules, which made the degradation feel broad and inconsistent. (anthropic.com) ### Why was the reasoning-effort change such a big deal? Because effort is basically the think-longer knob. Anthropic says high effort produced better output but sometimes made the UI feel frozen, so it lowered the default. Turns out that was the wrong tradeoff for many serious users, especially on coding tasks where a slower but more coherent answer is worth it. Anthropic reverted that change on April 7 after users pushed back. (anthropic.com) ### What was the “forgetful Claude” bug? The March 26 change was supposed to clear older thinking once when a session had been idle for more than an hour. A bug made that cleanup keep happening on every turn afterward. That made Claude seem forgetful and repetitive — exactly the kind of behavior people were complaining about. Anthropic says it fixed that on April 10. ### So where do the workarounds come in? (anthropic.com) Users filled the gap before Anthropic published its explanation. Some leaned harder on `CLAUDE.md`, which Claude Code officially supports as a persistent instruction file across surfaces. Others started forcing more explicit planning — asking Claude to outline steps first, keep a running plan, or restate constraints before editing code. Those are not magic hacks. They are ways of re-injecting structure when the default behavior feels less deliberate. ### Is a public `CLAUDE.md` actually a real thing? Yes. Anthropic’s docs explicitly say Claude Code surfaces share `CLAUDE.md` files, settings, and MCP servers. Anthropic’s own quickstarts repo also includes a `CLAUDE.md` file with project instructions, which tells you this is a first-class pattern, not a community superstition. ### Did Anthropic admit this changed user experience? Yes — pretty directly. It said the experience “isn’t what users should expect” and reset usage limits for subscribers as of April 23. (code.claude.com) That matters because it turns a vague social-media complaint cycle into a concrete product failure with dates, causes, and reversions. ### Bottom line? The interesting part is not that users invented rituals around Claude. (code.claude.com) Power users do that with every model. The interesting part is that Anthropic’s own postmortem now lines up with the shape of those complaints: less thinking, worse memory, weaker coding. The folk fixes spread because they were compensating for real regressions — and now we know which ones. (anthropic.com)