Anthropic cut 80% of Claude Code's system prompt, no quality loss
The six shifts that swap rigid rules for judgment, a claude doctor pass to slim your own agent prompts, and why some model endpoints are quietly getting dumber.
Copy markdown
80% gone, and the evals didn't budge
Anthropic pulled more than 80% of Claude Code's system prompt for its Claude 5-generation models (Opus 5, Fable 5) with no measurable loss on internal coding evals. The lesson for anyone shipping agents: most of your prompt is defensive cruft written for weaker models, and it's now costing you tokens and latency for nothing.
The six shifts to steal
The published playbook: rules to judgment, examples to interface design, upfront context to progressive disclosure via skills, deduped instructions across prompt and tools, auto-memory over hand-kept CLAUDE.md files, and rubrics or tests instead of verbose markdown. One directive went from “never write multi-paragraph docstrings” to “match the surrounding code's comment density and idiom.”
Audit your own prompt bloat this weekend
Point the checklist at your agents: strip rules your current model no longer needs, collapse duplicated instructions, and move static preamble into skills that load on demand. Builders are pairing Claude Code's /doctor check with the six-shift rubric to find what to cut — expect lower cost and latency at the same quality.
Elsewhere: are your models quietly getting dumber?
A widely-shared builder writeup argues some providers are silently serving quantized or smaller-routed variants behind unchanged API names to save compute, degrading output with no changelog. The defensive move: pin model versions where the API allows it and run a small canary eval so a stealth downgrade trips an alarm before your users notice.