Anthropic cut 80 percent of Claude Code's system prompt

Anthropic deleted more than 80% of Claude Code's system prompt for its Claude 5 generation models. The company reported no measurable loss on coding evaluations, per its post on context engineering, published on Opus 5's launch day.

For anyone maintaining a CLAUDE.md or AGENTS.md file, that means most of what's in there was built to compensate for a model that's already retired.

The more disruptive claim is about examples. Anthropic says giving the new models few-shot examples "constrains them to a certain exploration space," and recommends designing expressive tool interfaces instead.

Few-shotting has been standard practice since GPT-3. Anthropic is telling people to stop.

Anthropic's CLAUDE.md advice: keep it lightweight, say what the repo is for, and spend most tokens on codebase gotchas, not hard behavioral rules. Their example swap replaces "Never write multi-paragraph docstrings" with guidance to match the surrounding code's idiom and comment density.

paddo.dev corroborated the 80% figure with a cost number. One build dropped from 470k tokens ($1.29) on Opus 4.8 to 179k tokens ($0.33) on Opus 5. That's a 62% token cut and a 74% cost cut on identical work.

paddo.dev also audited 74 real CLAUDE.md files hunting for the "verify your work" instructions Anthropic says to cut, and found zero genuine self-check imperatives. Every "verify" hit was domain vocabulary, JWT checks, HMAC signatures, or an external command like running flutter analyze. A blind grep for that word would have broken all 74 repos.

My own project rules file has eighteen months of "Claude keeps doing X, stop it" entries. I couldn't tell you which ones a Claude 5 model still needs.

paddo.dev's fix: split permanent repo facts from perishable model patches, and review the second pile at every model release. I'm adopting it. Don't grep your rules file, read it.