Why Claude Code Cut 80% of System Prompts Overnight

Background and Trigger

When Anthropic released Claude Opus 5, the new generation of strong reasoning models, the team quickly realized that the existing prompt‑engineering paradigm—heavy, hand‑crafted system prompts, extensive CLAUDE.md files, and numerous Skills—was becoming counter‑productive. They observed that deleting more than 80% of the Claude Code system prompts did not cause any measurable degradation in coding‑related benchmarks.

Core Insight: Stronger Models Need Less "Nanny‑Style" Guidance

The team summarized the finding as: the stronger the model, the less it needs verbose, rule‑based instructions. Long‑hand rules waste tokens and can even hurt performance. Instead, a high‑performing team keeps the total prompt size under 60 lines (never exceeding 300) and moves task‑specific context into Skills.

From Rules to Judgment: Old vs. New System Prompts

Old approach: System prompts contained strict directives such as “never write multi‑line docstrings” and “always omit comments unless explicitly requested.” These rules sometimes conflicted with user preferences or complex code that genuinely required detailed comments.

Write code with no comments by default. Never write multi‑line docstrings or comment blocks—at most one short line. Unless the user explicitly asks, do not create planning, decision, or analysis documents.

Because earlier models needed explicit guardrails, such constraints were necessary. However, they also produced contradictory instructions within the same request, forcing the model to repeatedly weigh conflicting signals.

New approach: The revised system prompt is concise, e.g., “Generated code should match the surrounding code’s comment density, naming style, and idioms.” This lets the model use its own judgment based on surrounding context and its improved reasoning ability.

Generated code should read like the surrounding code: match its comment density, naming conventions, and idiomatic style.

Best Practices for Modern Context Engineering

System Prompt: Keep it tightly bound to the product context; avoid over‑specifying rules.

CLAUDE.md: Keep lightweight—briefly describe the repository’s purpose and allocate most tokens to actual code “pitfalls.” Use progressive disclosure for longer guidance.

Skills: Treat as lightweight guides that the model can consult on demand; avoid over‑constraining unless the domain is critical.

Reference Materials: Prefer code‑based artifacts (e.g., test suites, HTML artifacts) over textual descriptions; they provide high‑fidelity instructions.

Progressive Disclosure: Split large instructions into multiple files or Skills, loading them only when needed. Tools can use a “delayed‑load” pattern where the agent first discovers a tool definition via ToolSearch before invoking it.

Practical Workflow

When assembling the context for a Claude Code session, developers should:

Start with a minimal system prompt that defines the high‑level role.

Reference concise CLAUDE.md for repository‑wide conventions.

Expose task‑specific guidance through Skills, using progressive disclosure to keep each Skill focused.

Attach reference materials (e.g., spec files, test cases) via @ mentions so Claude can fetch deep details on demand.

Run the /doctor command (Claude Doctor) to automatically trim and balance the size of system prompts, Skills, and CLAUDE.md.

Conclusion

The shift from exhaustive, rule‑heavy prompts to a lean, architecture‑level context reflects the growing capabilities of large language models. By embracing minimal system prompts, progressive disclosure, and richer reference artifacts, developers can maintain or improve performance while reducing token consumption and complexity.