Anthropic explains three separate Claude Code bugs between March 4 and April 20 that degraded quality: reduced reasoning effort, dropped thinking history, and a verbosity system prompt.
Key Takeaways
March 4: reasoning effort silently dropped from high to medium for Sonnet 4.6 and Opus 4.6 to cut latency; reverted April 7 after user complaints.
March 26 caching bug cleared thinking history every turn instead of once, causing forgetfulness, repetition, and faster usage limit drain.
April 16 system prompt capped tool-call output at 25 words; ablations later found a 3% coding quality drop across Opus 4.6 and 4.7.
Three changes hit different traffic slices on different dates, making aggregate degradation look like normal variation and delaying diagnosis by weeks.
Going forward: xhigh effort is now the default for Opus 4.7, per-model evals run before every system prompt change, and usage limits are reset.
Hacker News Comment Review
Consensus: session state changes, context management behavior, and effort-level defaults were all modified without documentation; users discovered problems through degraded output, not changelogs.
Commenters pushed back on process: experimental gates, soak periods, or gradual rollouts for system prompt changes could have caught each regression earlier.
The 25-words-between-tool-calls instruction became a meme; one commenter called it “Claude caveman,” capturing frustration that a single prompt line tanked coding quality across all affected models.
Notable Comments
@everdrive: Reports Claude narrating its own internal prompt-injection defenses aloud in responses, a distinct issue not covered by any of the three disclosed bugs.
@natdempk: Session-resumption thinking truncation was undisclosed; “I don’t think that’s even documented behavior anywhere.”
