English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Claude 5 Context Engineering: Anthropic Cut 80% of the System Prompt with No Measurable Performance Drop — They Call It 'Unhobbling'

Forum topic · 小凯 · 2026-07-25

Summary

Anthropic engineer Thariq Shihipar published a post on 'the new rules of context engineering for Claude 5 generation models,' reporting that over 80% of the Claude Code system prompt was removed with no measurable decline on coding benchmarks for Opus 5 and Fable 5. This post summarizes the three core shifts: replacing hard rules with model judgment (e.g., matching a codebase's comment conventions instead of banning comments), replacing exhaustive tool examples with better interface design (expressive parameters and clear state enums), and replacing upfront instructions with progressive disclosure via the skill system and deferred tool loading through ToolSearch. Additional changes: deduplicated instructions, automatic memory replacing CLAUDE.md as primary memory, richer plan-mode artifacts, and a new /doctor command that trims CLAUDE.md and skill files. The concept 'Unhobbling' — acknowledging the team had been constraining the model with worst-case guardrails — suggests agent prompts should be slimmed down. Caveats: evidence covers coding evaluations only, and results apply to Opus 5 and Fable 5, not necessarily other model versions.

Anthropic's engineering team today published a long-form article on the new rules of context engineering for Claude 5 generation models. In it, Thariq Shihipar states directly that they removed more than 80% of the Claude Code system prompt and ran coding evaluations on Opus 5 and Fable 5 with no measurable performance decline.

That number may look like marketing at first glance, but the concrete accompanying changes are genuinely useful. I picked out three pairs of before/after contrasts worth noting:

1. Rules → Judgment

The old system prompt contained hard constraints 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." The new guidance is instead: match the comment density, naming, and conventions of the surrounding codebase. After mining large amounts of internal logs, the team found these strong constraints were sometimes wrong for certain prompts — users may genuinely want documentation, and some complex code legitimately needs multi-line comments. Newer models can make this call from context; they don't need rigid rules.

2. Examples → Interfaces

Tool descriptions used to be stuffed with many examples teaching Claude how to use them. The result: "boxing Claude into a particular exploration space." The new approach puts the effort into tool interface design — whether the parameters are expressive enough, whether state enumerations are clear. For example, the Todo tool's state enum (pending / in_progress / completed) itself tells Claude how to use it.

3. All-Upfront → Progressive Disclosure

Previously, all code review and validation details were packed into the system prompt. Claude Code now uses the skill system — validation lives in a standalone skill that Claude invokes on demand. Tools also use deferred loading: definitions must first be fetched via ToolSearch. This lets Claude Code ship more tools without crowding the context window.

Other Changes

  • Duplicated instructions removed — clear tool descriptions are enough.
  • CLAUDE.md is no longer the primary memory source — the model now automatically saves relevant memories itself.
  • Plan mode is no longer limited to markdown files; it can reference richer artifacts such as HTML, test suites, or rubrics.
  • The /doctor Command

    There's also a new command: /doctor (the original article writes "claude doctor," but Claude Code's command convention is /doctor). It automatically trims your CLAUDE.md and skill files. Anthropic ran it against their own internal codebase.

    Why "Unhobbling" Matters

    What interested me most after reading is the term "Unhobbling" — internal team jargon meaning "we've been hobbling Claude all along." As the models got stronger, guardrails originally designed to prevent worst-case outcomes became constraints the model "must think through carefully before deciding." Removing them actually made it more accurate. If this holds in external products too, it means agent system prompts and CLAUDE.md files everywhere are due for a round of slimming.

    Caveats

  • The methodology in the article is evaluated only on coding tasks. Whether it transfers to other domains (writing, design, customer support) is not backed by Anthropic data.
  • The "80% removed, no score drop" result applies to Opus 5 and Fable 5 specifically — it cannot be directly extrapolated to Sonnet or other model versions.
Original article

Tags

#anthropic#claude#context-engineering#claude-code#system-prompts#ai-agents#prompt-engineering#llm

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178447090