Anthropic Deleted 80% of Its System Prompts and Coding Scores Didn't Budge
Teams pouring effort into elaborate system prompts and CLAUDE.md files are likely wasting tokens and attention. The shift to progressive disclosure and interface-level constraints means less prompt maintenance and fewer conflicting instructions, but it also demands rethinking how rules are enforced—through sandboxes, hooks, and code rather than prose.
Anthropic's context engineering team found that newer Claude models perform just as well on coding benchmarks after removing more than 80% of the system prompt. The reason: strong models have a limited attention budget, and every extra rule forces the model to spend capacity judging whether it applies or resolving conflicts with other instructions. The old approach of hard-coding behaviors—no comments, no docstrings, no planning docs—often backfired by blocking necessary documentation on complex tasks.
The replacement is a single style rule: match the surrounding code's comment density, naming, and idioms. Tool constraints move into parameter design rather than prose instructions. Long-lived project knowledge splits across CLAUDE.md for mandatory rules, auto memory for discovered patterns, and Skills loaded on demand. References expand beyond Markdown specs to include runnable HTML prototypes, existing code, test suites, and scoring rubrics.
Anthropic draws a hard boundary: code style and documentation habits can be left to model judgment, but permissions, compliance, financial operations, and irreversible actions still require strict, enforced constraints outside the context layer.
The 80% figure is less about prompt engineering skill and more about model capability: older models needed rules to compensate for poor judgment, while newer models suffer when rules override their ability to read the codebase directly.
Conflicting instructions in context are a hidden failure mode. When a system prompt, a Skill, and a user request all say different things about comments, the model spends attention arbitrating rather than coding.
Moving constraints from prose into tool interfaces is a design principle that applies beyond AI: well-shaped parameters communicate intent more reliably than natural-language instructions ever can.
Progressive disclosure treats context like a cache hierarchy—keep the working set small, fetch detailed instructions only on a cache miss. This is a sharp departure from the 'employee handbook' model most teams still use.
The distinction between CLAUDE.md as mandatory rules and auto memory as discovered patterns creates a feedback loop where only recurring, team-relevant lessons get promoted to permanent context.