Claude 5 changed the rules of context engineering: over 80% of Claude Code’s system prompt is gone
Anthropic removed over 80% of Claude Code’s system prompt with zero loss on coding evals: the new rules of context engineering for Claude 5 generation models.

Luca Di Domenico
· 2 min read

In short
With the Claude 5 generation (Opus 5, Fable 5) Anthropic removed over 80% of Claude Code’s system prompt with no measurable loss on coding evaluations, as documented in Anthropic’s post of July 24, 2026. The context engineering paradigm flips: model judgment instead of rigid rules, progressive disclosure instead of loading all the context upfront, expressive tools and clear interfaces instead of detailed examples. In Claude Code, the /doctor command analyzes and trims skills and CLAUDE.md following these best practices.
🚀 Claude 5 changed the rules of context engineering
The old golden rule of context engineering
Until a few months ago the golden rule was clear:
“Put ALL the rules in the context, always. Otherwise the model won’t see them.”
Endless system prompts, CLAUDE.md files packed with instructions, examples on top of examples, repetitions...
Over 80% of Claude Code’s system prompt removed
We literally had to “rein in” the model to avoid disasters. With the new generation of Claude models (Opus 5, Fable 5…) Anthropic did something wild:
they removed over 80% of Claude Code’s system prompt → zero performance loss on coding evaluations.
The figure comes from the post “The new rules of context engineering for Claude 5 generation models”, published by Anthropic on July 24, 2026 and written by Thariq Shihipar.
The new paradigm: judgment, progressive disclosure, expressive tools
The new paradigm is completely different:
Before → Give rigid rules
Now → Let it use its judgment
Before → Put everything upfront
Now → Progressive disclosure (load only what’s needed, when it’s needed)
Before → Give detailed examples
Now → Design clear interfaces and expressive tools
In practice: you no longer need to force every piece of information into the context.
You need to give the model easy, smart ways to find what it needs on its own (on-demand skills, deferred tools, file trees, auto-memory, rich references…).
The model has become good enough that it no longer needs to be “controlled” obsessively.
It needs room and good entry points.
The /doctor command in Claude Code
đź’ˇ Practical tip: if you use Claude Code, just type /doctor.
It automatically analyzes and “trims” skills and CLAUDE.md, applying exactly these new best practices.
The lesson for anyone building agents
This is an important lesson not just for Claude Code users, but for anyone building agents or systems on the 5 line of Anthropic’s models.
Fewer rules. More trust. Better context design.
Are you already applying this approach?
Sources
- The new rules of context engineering for Claude 5 generation models — Thariq Shihipar (Anthropic), July 24, 2026: over 80% of Claude Code’s system prompt removed for models like Opus 5 and Fable 5 with no measurable loss on coding evaluations; the /doctor command trims skills and CLAUDE.md
- Anthropic Cut 80% of Claude Code’s System Prompt. Here’s Why That Matters for Your Agents — Techstrong.ai, July 27, 2026: independent confirmation of the 80%+ system prompt reduction for Opus 5 and Fable 5 and of the /doctor command rollout in Claude Code
Want to know what this means for your product?
If you're building software with AI and want a second opinion on how you're working, tell me about your project.
Related articles
Keep reading
Fable 5 found the cause of a year-long revenue drop with an SEO audit
Read: what Fable 5 found with an SEO audit