What did Anthropic actually change in Claude Code's system prompt?

Anthropic removed more than 80% of Claude Code's system prompt for Claude 5 generation models with no measurable drop on coding evaluations. Prescriptive guardrails such as 'default to writing no comments' were replaced with judgment-based guidance like 'match the surrounding code'. Explicit examples in tool descriptions and repeated instructions across system prompt and tools were also cut.

How do I apply Anthropic's Claude 5 context-engineering rules to my own setup?

Anthropic's playbook has six moves: enable judgment over rigid rules, design expressive tool interfaces instead of adding examples, use progressive disclosure through skills and deferred-loading tools, remove repeated instructions, trust auto-memory rather than manual CLAUDE.md edits, and provide rich references such as test suites and HTML artifacts as specifications.

What does the /doctor command do in Claude Code?

The /doctor command inside Claude Code (or `claude doctor` from your shell) audits your installation, settings, extensions, skills, CLAUDE.md, and context usage, and proposes fixes you can confirm. The July 24 post repositions it as the primary way to rightsize a Claude 4-era setup for Opus 5 and Fable 5 rather than only a health check.

Do the old Claude 4 context-engineering rules still work on Claude 5?

Old Claude 4 rules still work but Anthropic's post argues they leave performance on the table for Claude 5 models. Heavy guardrails, in-line examples, and instruction repetition were needed to keep Claude 4 aligned; Claude Opus 5 and Fable 5 handle those cases from judgment and lightweight guidance, so keeping the old scaffolding wastes tokens and can constrain exploration.

Does this only apply to Claude Code, or to any Claude 5 agent?

Thariq Shihipar frames the rules as general context engineering for the Claude 5 generation, using Claude Code as the worked example. Third-party agents built on Opus 5 or Fable 5 with the Anthropic SDK should see similar gains from progressive disclosure, expressive tool schemas, and lean system prompts; the /doctor command itself is specific to Claude Code.

Anthropic · 2026-07-24 · major

Anthropic's new context-engineering rules — 80% less system prompt on Claude 5

Anthropic cut over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with no measurable performance loss. A new best-practices playbook plus a repurposed /doctor command help users simplify their own skills and CLAUDE.md files.

Anthropic blog banner for 'The new rules of context engineering for Claude 5 generation models'

Anthropic just published the playbook for building Claude 5 agents — start by deleting 80% of your system prompt.

Quick facts

Author	Thariq Shihipar (Anthropic MTS, Claude Code)
Published	July 24, 2026
Applies to	Claude Opus 5, Claude Fable 5
Headline change	80%+ of Claude Code's system prompt removed with no eval loss
Tool	/doctor and `claude doctor` audit skills and CLAUDE.md
Old rule	Prescriptive guardrails and repeated instructions
New rule	Judgment over rules, progressive disclosure, tool interfaces over examples

What is it?

Anthropic engineer Thariq Shihipar's July 24 blog post lays out the new context-engineering rules for Claude Opus 5 and Claude Fable 5. The team cut more than 80% of Claude Code's own system prompt for these models with no measurable drop on coding evals, and the article turns that experience into concrete guidance for anyone building on the Claude 5 generation.

How does it work?

Claude 5 models reward judgment, not scaffolding. The playbook swaps prescriptive rules ('never write multi-paragraph docstrings') for terse guidance ('match the surrounding code'), replaces in-line tool examples with expressive tool interfaces, uses progressive disclosure through skills and deferred-loading tools instead of front-loading everything, and drops instruction repetition across the system prompt and tool descriptions. The /doctor command in Claude Code audits an existing setup against these rules and proposes fixes.

Why does it matter?

Prompts written for Claude Opus 4.x carry dead weight on Claude 5 — Anthropic's own experience suggests most Claude 4-era CLAUDE.md files, skills, and system prompts can be shrunk dramatically without losing capability. Following this playbook cuts token spend on every request, speeds up cold starts, and unlocks the exploratory behavior the newer models were tuned for. It is a direct answer to the 'my prompts got worse on Opus 5' complaints of the last week.

Who is it for?

Claude Code users, Anthropic SDK developers, agent builders migrating from Claude 4 to Opus 5 or Fable 5

Frequently asked questions

What did Anthropic actually change in Claude Code's system prompt?: Anthropic removed more than 80% of Claude Code's system prompt for Claude 5 generation models with no measurable drop on coding evaluations. Prescriptive guardrails such as 'default to writing no comments' were replaced with judgment-based guidance like 'match the surrounding code'. Explicit examples in tool descriptions and repeated instructions across system prompt and tools were also cut.
How do I apply Anthropic's Claude 5 context-engineering rules to my own setup?: Anthropic's playbook has six moves: enable judgment over rigid rules, design expressive tool interfaces instead of adding examples, use progressive disclosure through skills and deferred-loading tools, remove repeated instructions, trust auto-memory rather than manual CLAUDE.md edits, and provide rich references such as test suites and HTML artifacts as specifications.
What does the /doctor command do in Claude Code?: The /doctor command inside Claude Code (or `claude doctor` from your shell) audits your installation, settings, extensions, skills, CLAUDE.md, and context usage, and proposes fixes you can confirm. The July 24 post repositions it as the primary way to rightsize a Claude 4-era setup for Opus 5 and Fable 5 rather than only a health check.
Do the old Claude 4 context-engineering rules still work on Claude 5?: Old Claude 4 rules still work but Anthropic's post argues they leave performance on the table for Claude 5 models. Heavy guardrails, in-line examples, and instruction repetition were needed to keep Claude 4 aligned; Claude Opus 5 and Fable 5 handle those cases from judgment and lightweight guidance, so keeping the old scaffolding wastes tokens and can constrain exploration.
Does this only apply to Claude Code, or to any Claude 5 agent?: Thariq Shihipar frames the rules as general context engineering for the Claude 5 generation, using Claude Code as the worked example. Third-party agents built on Opus 5 or Fable 5 with the Anthropic SDK should see similar gains from progressive disclosure, expressive tool schemas, and lean system prompts; the /doctor command itself is specific to Claude Code.

Try it

claude doctor