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

Anthropic announced that it stripped over 80% of the system prompts used by Claude Code for Claude 5, and surprisingly the code‑evaluation scores barely changed. The author, who follows the original Chinese post, explores why the massive rule set that once helped weaker models now makes Claude 5 appear slower and less effective.

Too Many Rules Make Claude Appear Dumber

Many users let CLAUDE.md grow with rules such as “never add comments”, “only one line of docstring”, or “never modify code outside the task scope”. Over time the file fills with “must”, “must not”, and “never” statements that even the author can’t recall why they were added.

Anthropic’s internal logs show a conflict: users want documentation in the right places, but the system prompt says “don’t create analysis docs”. The model must decide which instruction to obey, and the cluttered context forces it to sift through historical patches to find the relevant rule.

From Rigid Constraints to Style Matching

Old Claude Code rules forced strict behavior, for example:

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。

These rules ignore project‑specific coding styles. The new prompt replaces them with a more flexible guideline:

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。

This shift lets Claude first look at the existing codebase, then decide how to act, similar to how a mentor would give a newcomer autonomy after they’ve learned the basics.

Context Engineering vs. Prompt Length

The author clarifies that “prompt” is only the user’s immediate instruction; the model’s full context also includes system prompts, CLAUDE.md, Skills, memory, and temporary file references. Traditional “context engineering” tries to cram as much information as possible, but Claude 5 often doesn’t need all that data for simple tasks like changing a button color.

Anthropic now uses progressive disclosure : only the information needed for the current step is loaded, while larger specifications (code‑review Skill, deployment Skill, etc.) are fetched on demand. This reduces token usage, avoids irrelevant constraints, and mirrors modular loading in software development.

Other Subtle Changes

Redundant instructions have been removed. Previously the same rule might appear in CLAUDE.md, a Skill, and a tool description, leading to inconsistencies.

Auto‑memory replaces the manual # tags that forced users to write persistent notes in CLAUDE.md. Now only project‑specific knowledge (e.g., legacy API restrictions) stays in the file.

Specification documents can now be concrete artifacts such as test suites, reference functions, or HTML prototypes instead of long markdown descriptions.

Impact on Token Usage and Performance

OpenAI’s GPT‑5.6 official prompt guide also recommends shorter prompts. Their internal coding‑agent evaluation shows that removing duplicate instructions and simplifying tool descriptions improves scores by 10‑15%, cuts token usage by 41‑66%, and reduces cost by 33‑67%.

When Skills Become Overkill

For developers using the Superpowers Skill with GPT‑5.6 Sol, the author observed that the mandatory “using‑superpowers” rule forces the model to run a full brainstorming, design, and verification workflow even for trivial changes, inflating token consumption and slowing execution.

The lesson is that as models become stronger, many heavyweight Skills that merely re‑specify steps become unnecessary and may hinder performance. Skills should add genuine new capabilities or knowledge, not just restate what the model can already decide.

Practical Recommendations

Audit your CLAUDE.md: remove rules the model can infer from the codebase (e.g., language choice, comment style).

Replace absolute directives (“never add comments”) with conditional guidelines that respect existing project conventions.

Modularize rarely used processes into separate Skills (e.g., Review Skill, Deploy Skill) and load them only when needed.

Use the /doctor command (or claude doctor) to detect overly heavy rules, but still review the suggestions manually.

Ultimately, as AI models grow more capable, the focus shifts from adding more prompts and Skills to trimming unnecessary constraints—doing “less” becomes the new way to get better results.

Reference: https://x.com/trq212/status/2080710971228918066