Anthropic cut 80% of Claude Code's system prompt for its newest models

Anthropic says it stripped more than 80% of Claude Code’s system prompt when targeting its latest models—Opus 5 and Fable 5—without measurable regression on coding benchmarks. The core lesson is that older prompting habits were built to prevent worst-case behavior in weaker models, and those guardrails now actively hobble models capable of reasoning through ambiguity on their own. Reading their own usage transcripts, the team found the accumulated layers of system prompt, skills, and CLAUDE.md files were issuing contradictory orders (for instance, ‘leave documentation as appropriate’ colliding with ‘DO NOT add comments’), forcing the model to burn effort reconciling conflicts instead of just inferring intent.

The post reframes several once-standard practices as outdated myths. Rigid rules give way to model judgment (e.g., ‘match the surrounding code’s comment density’ instead of blanket comment bans); hard-coded tool examples give way to well-designed, self-describing interfaces, since examples now narrow the model’s exploration space; and dumping everything upfront gives way to progressive disclosure, where context—including deferred-loading tools fetched via search—is pulled in only when needed. Repetition and redundant instructions are replaced by clear tool descriptions, manual CLAUDE.md memory notes by automatic memory, and plain markdown specs by richer references like HTML artifacts, test suites, code to port, and rubrics that verifier agents can check work against.

For practitioners building their own agents, the guidance is to invest heavily in the system prompt (which defines the product and rarely changes), keep CLAUDE.md lightweight and focused on codebase gotchas rather than obvious facts, and lean on progressive disclosure through a tree of context files loaded at the right moment. Anthropic has baked these recommendations into a /doctor command in Claude Code that helps users right-size their skills and CLAUDE.md files. The broader significance: as model capability climbs, the highest-leverage prompt engineering increasingly means removing constraints, not adding them.