Anthropic's engineering team has published a long-form post titled "The New Rules of Context Engineering for Claude 5 Generation Models." Author Thariq Shihipar states directly that they deleted more than 80% of Claude Code's system prompts and ran coding evaluations on Opus 5 and Fable 5 — with no measurable performance drop.
That number may sound like marketing at first glance, but the concrete actions behind it are genuinely useful. Here are three key contrasts distilled from the original post:
1. Rules → Judgment
The old system prompt contained hard rules like "don't write comments in code by default," "don't write multi-paragraph docstrings," and "don't create plan documents unless the user asks." These are now replaced with: match the comment density, naming, and conventions of the surrounding codebase.
The team analyzed large volumes of logs and found these rigid constraints were sometimes wrong for certain prompts — users may actually want documentation, and complex code may genuinely need multi-line comments. Newer models can use context to make that judgment call without hard rules.
2. Examples → Interfaces
Tool descriptions used to be stuffed with examples teaching Claude how to behave. The result: it "boxed Claude into a narrow exploration space." The new approach invests in tool interface design — are the parameters expressive enough, are state enums clear? For example, the Todo tool's state enum (pending/in_progress/completed) itself communicates how to use it.
3. Front-loading → Progressive Disclosure
Instead of packing all code review and verification details into the system prompt, Claude Code now uses a skills system — verification is a standalone skill Claude invokes on demand. Tools also use "deferred loading": definitions must first be pulled via ToolSearch. This lets Claude Code carry more tools without consuming context.
Other Changes
- Duplicate instructions were deleted — clear tool descriptions are sufficient.
- CLAUDE.md is no longer the primary memory source; the model automatically saves relevant memories itself.
- Plan mode is no longer limited to markdown files — it can reference HTML artifacts, test suites, rubrics, and other richer artifacts.
- A new command,
/doctor(the original article says "claude doctor," but Claude Code's command convention is/doctor), automatically trims your CLAUDE.md and skill files. Anthropic ran it against their own internal codebase. - The methodology is validated only on coding evaluations. Whether it transfers to writing, design, or customer service is unsupported by Anthropic's data.
- The "delete 80%, no drop" result applies to Opus 5 and Fable 5 — it cannot be directly extrapolated to Sonnet or other model versions.
The Meaning of "Unhobbling"
The term that stands out most is "Unhobbling" — internal team jargon meaning "we've been hobbling (constraining) Claude all along." As models got stronger, guardrails designed to prevent worst-case outcomes became constraints the model "must think through carefully before deciding." Removing them actually improved accuracy. If this holds for external products, it suggests agent system prompts and CLAUDE.md files everywhere deserve a round of pruning.