Why Anthropic Cut 80% of Claude Code Prompts Without Dropping Performance

Recently Anthropic announced that they trimmed over 80% of the system prompts used by Claude Code for Claude 5, and the code‑evaluation scores barely changed. This surprising result led the author to examine the underlying reasons.

Rule bloat made Claude appear less capable

Many projects accumulated long CLAUDE.md files full of absolute directives such as “must never add comments” or “never modify files outside the task scope”. Each time the model made a mistake, a new rule was added, resulting in a file saturated with “must”, “never”, and “always” statements that the model had to parse before acting.

Anthropic’s internal logs showed conflicts: users wanted documentation where appropriate, while the system prompt prohibited creating it. The model had to decide which rule to obey, and the growing context buried the truly important information under a mountain of historical patches.

From hard constraints to project‑aware guidelines

Older prompts forced Claude to follow rigid constraints (e.g., "default no comments", "only one line of block comment"). The new prompts replace these with relative standards: "write code that matches the surrounding style" and "maintain similar comment density and naming conventions". This shift lets Claude first inspect the project and then decide how to comply, rather than being forced into a single prescribed behavior.

Example of the old rule set:

默认不写注释。
不要写多段文档字符串。
多行注释最多只能写一行。

And the new, more flexible version:

写出的代码要像周围原有的代码。
保持相近的注释密度、命名方式和代码习惯。

Context engineering: less is more

The author clarifies the difference between the user’s prompt and the full context Claude receives, which includes the system prompt, CLAUDE.md, Skills, memory, and temporary file references. Modern practice, called “progressive disclosure”, only injects the pieces needed for the current task, loading additional Skills (e.g., code review, deployment) on demand. This reduces token consumption and prevents irrelevant rules from distracting the model.

By moving large instruction blocks into separate Skills and loading them only when required, Claude’s effective context becomes lighter, similar to on‑demand module loading in software development.

Other subtle changes

Duplicate instructions across CLAUDE.md, Skills, and tool descriptions were eliminated, preventing inconsistencies.

Auto‑memory now automatically records relevant project and user information, removing the need to manually embed such data in CLAUDE.md.

Specification formats expanded beyond text: tests, HTML prototypes, or reference functions can now serve as concrete specifications, reducing ambiguity.

Impact on tokens, cost, and speed

OpenAI’s GPT‑5.6 official guide also recommends shorter prompts. Internal coding‑agent benchmarks showed that after removing duplicate directives and simplifying tool descriptions, scores improved by roughly 10‑15%, token usage dropped 41‑66%, and costs fell 33‑67%.

Conversely, overly detailed Skills like Superpowers (which force brainstorming, design, planning, and sub‑agent creation even for trivial changes) can inflate token usage and slow execution when the model already knows how to handle the task.

Practical advice for cleaning CLAUDE.md

Delete information the model can infer from the codebase (framework choice, naming conventions, directory layout).

Retain only project‑specific knowledge that the model cannot discover (legacy API constraints, safety boundaries, special deployment quirks).

Replace absolute rules with conditional guidelines that reference the existing code style.

Split rarely used processes into separate Skills (e.g., Review Skill, Deploy Skill) instead of keeping everything in a monolithic CLAUDE.md.

Use the /doctor command (or claude doctor) to audit the current prompts and identify overly heavy sections.

Ultimately, as models become stronger, the focus shifts from adding more prompts to performing prompt reduction, ensuring that AI agents receive only the essential context they truly need.