Why Dropping 80% of System Prompts Improves Claude 5: New Context Engineering Rules

Context prompt reduction

Anthropic’s guide reports that for Claude Opus 5 and Claude Fable 5, more than 80 % of the Claude Code system prompt was removed, and coding benchmark scores showed no measurable difference.

Conflicting directives

Internal transcriptions show requests that simultaneously contain “leave documentation as appropriate” and “DO NOT add comments”. Older models required hard‑coded rule overrides; the newer model’s judgment makes those constraints unnecessary, so they were deleted.

Six paradigm reversals (Then → Now)

Rules → model judgment : Previously the system prompt forced “don’t write comments”. The new prompt contains only “Write code that reads like the surrounding code: match its comment density, naming, and idiom.”

Examples → interface design : Earlier examples locked the model to a narrow space. Replacing them with expressive tool parameters, e.g. an enum Todo with values pending, in_progress, completed, lets the model infer intended usage.

All upfront → progressive disclosure : Instead of embedding the entire review and verification flow in the system prompt, the new approach loads the appropriate context at execution time, using deferred loading and ToolSearch to fetch skills only when needed.

Repetition → concise tool description : Earlier the same instruction appeared in both the system prompt and the tool description; now it resides solely in the tool description.

Manual memory → automatic memory : Previously users added memory entries with a “#” hotkey in CLAUDE.md. The model now automatically saves memories related to the current work.

Simple spec → rich references : Simple markdown plans have been replaced by references to HTML artifacts, full test suites, functions from other codebases, and verifier agents invoked with rubrics to assess API design quality.

Four context carriers

system prompt

: Tightly bound to the product, defines the runtime environment. Claude Code users rarely modify it; custom agent harness developers invest heavily here. CLAUDE.md: Kept lightweight, describes repository purpose. Most tokens are spent on internal “gotchas”; obvious file‑system details are omitted, and similar instructions are split into reusable skills. skills: Light‑weight guides fetched on demand; long skills are broken into multiple files for hierarchical loading. References: Use @ to cite files, preferring code files (e.g., HTML mockups) over textual descriptions or screenshots for higher fidelity.

Full guide: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models