Anthropic Cut 80% of Claude Code's System Prompt and Coding Scores Didn't Move

Anthropic rewrote Claude Code’s system prompt for the Claude 5 generation of models and cut more than 80% of its instructions without hurting coding-eval performance. That’s the headline result behind what Anthropic is calling “context engineering” — the discipline of deciding what a model sees, and when, across system prompts, CLAUDE.md files, tool descriptions, and memory. The shift is philosophical as much as technical: trust model judgment instead of hardcoding rules, load information only when it’s needed instead of front-loading everything, and cut instructions that exist only to compensate for older, weaker models. Anthropic’s own example is a rigid ban on multi-line comments, replaced with the looser instruction to match the surrounding code’s comment density and idiom. They call this “unhobbling” — removing scaffolding a more capable model no longer needs.

That principle shows up elsewhere in Anthropic’s own tooling the same week. A companion post on verification loops describes the inverse move — not fewer instructions, but turning repeated manual checks into reusable, model-invoked “skills” so Claude closes its own feedback loop instead of waiting on a human to catch mistakes. And Datadog’s internal system, Temper, built on Claude Code, pushes the same logic one layer deeper: agents don’t generate application code directly, they produce verifiable specifications checked by symbolic reasoning and fault injection before anything executes. Read together, the pattern is consistent — as models get more capable, the engineering work shifts from constraining what the model can do to designing what it’s allowed to verify for itself.

For consulting engagements building or maintaining Claude-based deployments, the practical takeaway isn’t “write a shorter prompt.” Anthropic also shipped a /doctor command in Claude Code specifically to audit an existing context setup for redundant or stale instructions — a concrete signal that most production CLAUDE.md files and system prompts accumulated during the era of weaker models are now overdue for a prune, not an expansion.