April 23, 2026
Caveman prompts, intern vibes, coder rage
An update on recent Claude Code quality reports
Users say Claude “forgot how to code” as Anthropic blames three oopsies and hits reset
TLDR: Anthropic says three missteps made Claude Code feel dumber (less thinking, a memory-wipe bug, and anti-verbosity prompts), and it rolled them back and reset limits. Users are split between appreciating the honesty and feeling burned—memeing “caveman” and “intern” while questioning trust and premium pricing.
Anthropic just owned up to a month of “why is my robot dev acting weird?” chaos—and the comments section is on fire. The company says three back‑to‑back tweaks made Claude Code feel dumber: lowering default “thinking” from high to medium, a bug that wiped past thoughts every turn (so it kept repeating itself), and a new instruction to be less wordy that accidentally nerfed coding quality. Everything’s now rolled back (as of April 20) and subscribers got usage limits reset on April 23. The API wasn’t affected—but try telling that to users who felt gaslit by the UI.
The community’s strongest vibe? Trust took a hit. One user tallied the timeline like a courtroom exhibit, noting the app showed “high” even when the default secretly went medium. Another compared Claude to an “intern” who nods, then does nothing until you nag, while a fed-up commenter dropped the hammer: “Damage is done.” The anger isn’t just about bugs; pricing A/B tests this week poured gasoline on the bonfire.
There’s real debate: some say thanks for the transparency, others call it too little, too late for a premium product. The memes wrote themselves: “Claude caveman in the system prompt confirmed?” cracked one commenter, while another blamed the “frozen UI” for bringing Ice Age vibes to their sprint. Bottom line: Anthropic says it never intentionally degrades models, but the crowd wants receipts—and a future where the “smart mode” actually stays smart.
Key Points
- •Three separate product-layer issues affected Claude Code, the Claude Agent SDK, and Claude Cowork; the API and inference layer were unaffected.
- •Default reasoning effort was changed from high to medium on March 4 to reduce latency and reverted on April 7 after users reported lower intelligence (affecting Sonnet 4.6 and Opus 4.6).
- •A March 26 fix for idle sessions introduced a bug that cleared thinking every turn, causing forgetfulness; fixed on April 10 (affecting Sonnet 4.6 and Opus 4.6).
- •An April 16 system prompt to reduce verbosity hurt coding quality; reverted on April 20 (affecting Sonnet 4.6, Opus 4.6, and Opus 4.7).
- •All issues were resolved by April 20 (v2.1.116), and usage limits are being reset for all subscribers as of April 23.