Anthropic Cut 80% of Claude Code's System Prompt — and Nothing Broke

Anthropic has published a breakdown of how the principles of “context engineering” — the crafting of system prompts, CLAUDE.md files, skills, and memory — have changed for the new Claude 5 generation of models (Opus 5, Fable 5). The headline sounds almost provocative: more than 80% of the text was removed from Claude Code's system prompt, and on internal coding benchmarks this produced no noticeable drop in quality.

From Rules to Judgment

The author, Anthropic technical staff member Tariq Shihipar, admits that the model used to be simply overloaded with constraints. A single request could juggle contradictory instructions — “keep documentation where needed” right alongside “do NOT write comments.” Older models without such “guardrails” went wrong too often, and that was something you had to live with. The new ones handle these forks on their own. The rigid wording “by default, do not write comments, no multiline blocks” was replaced with a single phrase: write code in the style of its surroundings — with the same comment density and the same idioms.

Busted Myths

Shihipar lists the former “best practices” that have become myths:

• Tool-use examples now do more harm than good — they box the model into a narrow search space. Instead of examples, the advice is to think through the design of the tools themselves and their parameters.
• “Front-loading everything into the start of the context” has given way to progressive disclosure: code checking and code review were spun off into separate skills, while some tools are loaded lazily, through search, and take up no context until they are needed.
• Repeating instructions is no longer necessary: the description of how to use a tool has moved into the description of the tool itself.
• Manually saving to memory via the # hotkey has been replaced by automatic memory — the model itself decides what is worth remembering about the task and about the user.
• Simple specifications in the form of markdown are yielding to “rich references”: a specification can be a test suite, a function from another project, or an HTML artifact, while quality criteria are framed as rubrics for verifier agents.

What to Do With Your Own Context

The advice for those building their own agents boils down to asceticism. The CLAUDE.md file should be kept lightweight, spending tokens not on the obvious (the model will see that in the repository anyway) but on the codebase's “pitfalls.” Skills are lightweight guides, not a compendium of every conceivable rule; long ones are better split across files. As references, code in a language the model understands is preferable: an HTML mockup will yield a better result than a verbal description or a screenshot.

To help tidy things up, Anthropic has released the claude doctor command (/doctor inside Claude Code), which helps “right-size” skills and configuration files automatically.

Source: Hacker News