Cutting 80% of Claude’s System Prompts Still Yields Strong Performance

Anthropic’s official blog announced that for the new Claude Opus 5 and Fable 5 models they cut over 80% of the system prompts, and the coding evaluation results did not drop.

The author reflects that many earlier prompt rules were compensating for the limited judgment of older models; as model capabilities improve, overly detailed rules become constraints.

Two concepts from Anthropic’s context‑engineering series help explain this:

Attention Budget : Like human working memory, the model’s attention is limited; more context rules consume attention that could be used for actual work.

Marginal Diminishing Returns of Context : Adding more tokens reduces the model’s ability to precisely retrieve information, similar to overloading a person with a 100‑page document.

Anthropic Specific Changes

1. From Fixed Rules to Model Judgment

Old Claude Code system prompt forced a strict rule:

默认不写注释。永远不要写多行文档字符串或多行注释块，最多一行简短注释。除非用户要求，否则不要创建规划、决策或分析文档。

This “one‑size‑fits‑all” rule lacked flexibility. The new prompt is a single sentence:

写出来的代码要像周围的代码，匹配它的注释密度、命名方式和惯用法。

The model now adapts to the project's existing code style instead of being constrained by pre‑set rules.

2. From Examples to Well‑Designed Interfaces

Previously, a classic few‑shot technique gave the AI three examples of how to call a tool. Anthropic found that for the new model, examples limit exploration because the model already understands the tool better than the examples provide.

Instead, they suggest designing the tool’s interface clearly. For the Todo tool in Claude Code, the parameters are: pending, in_progress, completed. The model can infer usage without additional examples.

3. From All‑In to On‑Demand Loading

Earlier Claude Code prompts bundled extensive guidance on code review, validation, etc., which was rarely needed but critical when required, so everything was packed together.

Anthropic now splits these into independent Skills packages and loads them only when needed—a strategy they call Progressive Disclosure . Even tool definitions can be “lazy‑loaded” via ToolSearch when the model actually needs them.

4. From Manual Memory to Automatic Memory

Previously users had to press # to manually save important information to CLAUDE.md. The new model automatically saves relevant memories.

Consequently, CLAUDE.md should stay lightweight, focusing on project “gotchas” and special conventions that the model cannot infer from the codebase, while things like directory structure or dependency lists can be omitted.

5. Contradictory Instructions

Anthropic discovered that some system prompts said “add documentation when appropriate,” while Skills instructed “do not add comments,” creating a conflict. The model spends extra attention resolving such contradictions, so redundant rules were removed.

Practical Comparison

The author tested Cursor + Claude Opus 5 with two prompt styles to recreate a Cursor‑like web AI coding tool.

Long prompt (rule‑heavy) included many technical requirements and steps (image omitted).

Short prompt (5 lines) specified only the goal and research approach:

基于 VS Code 开源生态做一个类似 Cursor 的 Web AI 编程工具，
支持 Editor Window 和 Agents Window。
先用 Firecrawl 搜 Cursor 3 的产品设计，
再用 Context7 查 monaco-vscode-api 的 Web 集成方案，
调研完先出技术方案，确认后再写代码。

During the planning phase, the short prompt entered Plan mode and generated a complete Todo list, while the long prompt jumped straight into coding.

Code line count: short prompt produced 12984 lines vs. long prompt 8384 lines – about a third fewer for the long version.

Functionally, the short‑prompt version ran flawlessly: editor and agent windows switched, file tree loaded, CRUD operations worked, syntax highlighting, auto‑completion, git integration, theme switching, terminal, search, shortcuts, and agent‑mode code generation all succeeded.

The long‑prompt version showed many failures and missing features (screenshots omitted).

Both versions could call AI agents to generate a Snake game using DeepSeek V4 Flash, but the short‑prompt version displayed the reasoning chain and tool usage, offering a richer interaction.

The conclusion: long prompts constrain the model, limiting exploration, whereas short prompts give the model freedom to fill in missing functionality, leading to higher code quantity, quality, and success rate.

What We Should Do

When a model’s knowledge already covers a task’s common sense (e.g., writing a REST API that returns JSON), no prompt is needed. Only special, non‑standard requirements (e.g., mandatory snake_case naming) should be explicitly stated.

The code repository itself serves as a strong implicit prompt: project structure, style, and tests are read by the model. Over‑specifying rules wastes token budget and attention.

/doctor Command for Optimizing Project Rules

Claude Code includes a /doctor (or /checkup) command that scans Skills, CLAUDE.md, and other config files to detect redundant or unnecessary content.

It flags items like explicit directory structures or dependency lists that the model can read directly, suggests removing unused Skills, finds duplicate or conflicting rules, and recommends moving rarely‑used content into on‑demand Skills. Changes are presented as a report for user confirmation before execution.

The author’s own project test showed the command cleaned up some configuration but left CLAUDE.md largely unchanged, indicating the project’s rules were already healthy.

Final Thoughts

AI coding is evolving rapidly; what once seemed essential—extensive prompt engineering—can now become a hindrance. Anthropic’s closing remark summed it up: “Find the smallest, high‑signal token set to maximize the desired outcome.” In practice, less is more: many prompts can be discarded.