Anthropic Publishes Claude Code Post-Mortem — Three Separate Bugs Over Six Weeks, Not a 'Nerf'

Anthropic Engineering·high signal

Anthropic's April 23 post-mortem traced a month of user complaints to three distinct issues: a March 4 reasoning effort downgrade (high→medium), a March 26 cache-clearing bug that made Claude 'forgetful' every turn, and an April 16 system prompt change that hurt coding quality. All three were reverted by April 20 (v2.1.116). Boris Cherny, Claude Code's creator, posted additional details on Reddit (307 upvotes, 127 comments), confirming the API was never affected — only Claude Code, Agent SDK, and Cowork.