Jay E | RoboNuggets
24 min video
3 min read
Claude Code's 6 New Rules (2025 Update)
You just saved 21 min.
The big takeaway
Anthropic engineers revealed six major shifts in how to optimize Claude Code: let models use judgment instead of rigid rules, provide design interfaces over examples, use progressive disclosure to load context on-demand, simplify tool descriptions, embrace automatic memory, and replace markdown-only specs with richer HTML references. These changes reflect Claude's dramatically improved capabilities—Opus 5 scores 60/100 on intelligence benchmarks vs. 31/100 a year ago.
The Context Engineering Shift
Claude's Intelligence Leap
Claude's raw intelligence has nearly doubled in one year. Opus 4 scored 31% on the Artificial Analysis benchmark; Opus 5 now scores 60%, enabling Anthropic to remove over 80% of Claude Code's system prompt with zero performance loss.
Opus 4 (1 year ago)
31%
Opus 5 (now)
60%
Artificial Analysis Intelligence Benchmark
The ARMS Framework
Context engineering revolves around four elements: Applications (tools, MCPs, APIs), Routines (scheduled tasks), Memory (artifacts, reports), and Skills (SOPs and processes). Organizing these properly puts you ahead of 99% of Claude users.
1
Applications
Tools, MCPs, APIs
2
Routines
Scheduled tasks, crons
3
Memory
Artifacts, reports, docs
4
Skills
SOPs, processes, commands
ARMS Framework for Context Engineering
Rule 1: Let Claude Use Judgment, Not Just Rules
From Rigid Rules to Intelligent Discretion
Previously, Claude Code had strict rules like 'never write multi-paragraph docstrings' to prevent worst-case scenarios. Now, smarter models benefit from looser guidance—e.g., 'write code that reads like surrounding code, match its comment density'—allowing Claude to apply judgment contextually.
Old approach
Rigid, specific rules
New approach
Contextual judgment
System Prompt Philosophy Shift
Creative Skills Benefit Most
Skills designed for creative output (e.g., 'surprise me' for front-end design) perform better when Claude has freedom to iterate and explore rather than follow prescriptive rules. This generates unexpected, high-quality designs that rigid constraints would prevent.
Rule 2: Design Interfaces Over Examples
Examples Constrain, Interfaces Enable
Older models needed specific examples to understand tool usage. Newer models like Opus 5 and Fable 5 are constrained by examples—they explore less. Instead, provide a design interface (e.g., HTML brand book with color palettes, fonts, visual style) that guides without limiting exploration.
Old approach
Specific examples
New approach
Design interface system
Tool Guidance Strategy
Building a Design System Skill
Create an HTML brand book (e.g., /robo skill) that defines color palettes, fonts, voice, and visual style. Link it to a skill.md file. Claude then applies these guidelines consistently across all materials without being locked into specific examples.
Rule 3: Progressive Disclosure Over Upfront Context
Load Context When Needed, Not All at Once
Instead of stuffing Claude.md with every operational rule and practice, make it a router that points to specialized files. Load detailed context only when relevant to the current task. This reduces token waste on every session start and improves efficiency as workspace complexity grows.
1
Claude.md acts as central router
2
Points to department-specific files (content, product, community)
3
Sub-routers (e.g., content.md) guide to specific skills and references
4
Context loads progressively during session, not upfront
Progressive Disclosure Architecture
Token Cost Savings
A thin Claude.md router saves tokens on every session start. As you interact more with Claude, efficiency gains compound because you avoid re-loading a bloated system prompt. This matters significantly for high-volume users.
Sub-Router Example: Content.md
A content router file lists skills and file references for ideation, research, and production. Claude navigates to the right resources based on intent, rather than searching through a monolithic Claude.md.
Rule 4: Simpler Tool Descriptions
No More Repetition Across Layers
Older models suffered from context rot—they listened more to recent instructions than early ones, requiring repetition. Newer models are smarter and don't need tool descriptions duplicated in both system prompt and tool definition. Remove redundancy and save tokens.
Old approach
Duplicate instructions
New approach
Single, clear description
Tool Description Efficiency
Rule 5: Automatic Memory Over Manual Saving
Claude Now Auto-Saves Relevant Memories
Claude Code can automatically save memories relevant to your work. While manual saving (via # hotkey) is still useful for important learnings, the model now proactively captures and retains context from productive sessions.
The /calibrate Skill for Session Capture
Run /calibrate at the end of a session to review the conversation and automatically update skills, Claude.md, memory files, and prompt packs. This refines your operating system over time and helps Claude learn your preferences.
Rule 6: Richer References Beyond Markdown
HTML Artifacts as Living References
Move beyond markdown-only specs. HTML files (e.g., brand books, infographics) allow Claude to parse code while humans see visual rendering. This is especially powerful for design systems, color palettes, and complex specifications that markdown cannot convey effectively.
Old approach
Markdown files only
New approach
HTML + markdown hybrid
Reference Format Evolution
Visual Communication Advantage
HTML infographics and interactive references are easier to review than walls of markdown text. They're also more effective for communicating ideas to others and for investigating new skills or concepts.
Implementation: /doctor and /doctor-plus Skills
Built-in /doctor Skill
Claude Code's native /doctor command checks for broken installs, duplicate files, dead weight in skills and MCPs, slow hooks, and version updates. It reports findings and applies fixes to optimize your setup.
1
Check install integrity and path problems
2
Identify dead weight (unused skills, MCPs)
3
Find performance-draining hooks
4
Report version status
5
Apply recommended fixes
/doctor Skill Workflow
/doctor-plus for Advanced Optimization
An enhanced skill (available free) that runs all /doctor checks plus validates against the six new rules. It flags skills that violate new best practices (e.g., overly thick skill files that should be routers) and suggests optimizations.
Example Optimization: Last 30 Days Skill
A 2,090-line research skill flagged as too thick. Solution: refactor it into a router that points to specialized research sub-skills, reducing token overhead and aligning with progressive disclosure principles.
Worth quoting
"They actually removed over 80% of Claude Code's system prompt with no measurable loss."
— Jay E (summarizing Tariq's findings), at [3:02]
"It's actually better to just let your AI agent use a bit of judgment in order for you to get better results."
— Jay E, at [6:05]
"Your Claude.md becomes more powerful if you make it function as a router to your tree of files."
— Jay E, at [11:45]
Try this
Update Claude Code to the latest version to access /doctor and new features.
Audit your Claude.md: convert it from a monolithic reference into a router that points to department-specific files (content, product, community, etc.).
Create sub-routers (e.g., content.md) that guide Claude to relevant skills and files for specific intents.
Refactor overly thick skills (>1,000 lines) into routers with specialized sub-skills.
Build a design interface system (HTML brand book) and link it to a skill file (e.g., /robo) instead of providing specific design examples.
Replace rigid rules in skill descriptions with contextual guidance that allows Claude to use judgment.
Convert visual specifications from markdown to HTML artifacts (infographics, brand books) for better clarity.
Run /doctor-plus monthly to check your setup against the six new rules and optimize accordingly.
Set up a /calibrate routine at the end of productive sessions to auto-update skills, memory, and workflows.
Made with Glimpse by Wozart
glimpse.wozart.com/v/eyt2ldp6
Share this infographic
Read this infographic as text

Claude Code's 6 New Rules (2025 Update)

Summary of the video “Claude Code Just Changed Forever (6 NEW Rules by Anthropic Engineers) by Jay E | RoboNuggets.

Anthropic engineers revealed six major shifts in how to optimize Claude Code: let models use judgment instead of rigid rules, provide design interfaces over examples, use progressive disclosure to load context on-demand, simplify tool descriptions, embrace automatic memory, and replace markdown-only specs with richer HTML references. These changes reflect Claude's dramatically improved capabilities—Opus 5 scores 60/100 on intelligence benchmarks vs. 31/100 a year ago.

The Context Engineering Shift

Claude's Intelligence Leap

Claude's raw intelligence has nearly doubled in one year. Opus 4 scored 31% on the Artificial Analysis benchmark; Opus 5 now scores 60%, enabling Anthropic to remove over 80% of Claude Code's system prompt with zero performance loss.

The ARMS Framework

Context engineering revolves around four elements: Applications (tools, MCPs, APIs), Routines (scheduled tasks), Memory (artifacts, reports), and Skills (SOPs and processes). Organizing these properly puts you ahead of 99% of Claude users.

Rule 1: Let Claude Use Judgment, Not Just Rules

From Rigid Rules to Intelligent Discretion

Previously, Claude Code had strict rules like 'never write multi-paragraph docstrings' to prevent worst-case scenarios. Now, smarter models benefit from looser guidance—e.g., 'write code that reads like surrounding code, match its comment density'—allowing Claude to apply judgment contextually.

Creative Skills Benefit Most

Skills designed for creative output (e.g., 'surprise me' for front-end design) perform better when Claude has freedom to iterate and explore rather than follow prescriptive rules. This generates unexpected, high-quality designs that rigid constraints would prevent.

Rule 2: Design Interfaces Over Examples

Examples Constrain, Interfaces Enable

Older models needed specific examples to understand tool usage. Newer models like Opus 5 and Fable 5 are constrained by examples—they explore less. Instead, provide a design interface (e.g., HTML brand book with color palettes, fonts, visual style) that guides without limiting exploration.

Building a Design System Skill

Create an HTML brand book (e.g., /robo skill) that defines color palettes, fonts, voice, and visual style. Link it to a skill.md file. Claude then applies these guidelines consistently across all materials without being locked into specific examples.

Rule 3: Progressive Disclosure Over Upfront Context

Load Context When Needed, Not All at Once

Instead of stuffing Claude.md with every operational rule and practice, make it a router that points to specialized files. Load detailed context only when relevant to the current task. This reduces token waste on every session start and improves efficiency as workspace complexity grows.

Token Cost Savings

A thin Claude.md router saves tokens on every session start. As you interact more with Claude, efficiency gains compound because you avoid re-loading a bloated system prompt. This matters significantly for high-volume users.

Sub-Router Example: Content.md

A content router file lists skills and file references for ideation, research, and production. Claude navigates to the right resources based on intent, rather than searching through a monolithic Claude.md.

Rule 4: Simpler Tool Descriptions

No More Repetition Across Layers

Older models suffered from context rot—they listened more to recent instructions than early ones, requiring repetition. Newer models are smarter and don't need tool descriptions duplicated in both system prompt and tool definition. Remove redundancy and save tokens.

Rule 5: Automatic Memory Over Manual Saving

Claude Now Auto-Saves Relevant Memories

Claude Code can automatically save memories relevant to your work. While manual saving (via # hotkey) is still useful for important learnings, the model now proactively captures and retains context from productive sessions.

The /calibrate Skill for Session Capture

Run /calibrate at the end of a session to review the conversation and automatically update skills, Claude.md, memory files, and prompt packs. This refines your operating system over time and helps Claude learn your preferences.

Rule 6: Richer References Beyond Markdown

HTML Artifacts as Living References

Move beyond markdown-only specs. HTML files (e.g., brand books, infographics) allow Claude to parse code while humans see visual rendering. This is especially powerful for design systems, color palettes, and complex specifications that markdown cannot convey effectively.

Visual Communication Advantage

HTML infographics and interactive references are easier to review than walls of markdown text. They're also more effective for communicating ideas to others and for investigating new skills or concepts.

Implementation: /doctor and /doctor-plus Skills

Built-in /doctor Skill

Claude Code's native /doctor command checks for broken installs, duplicate files, dead weight in skills and MCPs, slow hooks, and version updates. It reports findings and applies fixes to optimize your setup.

/doctor-plus for Advanced Optimization

An enhanced skill (available free) that runs all /doctor checks plus validates against the six new rules. It flags skills that violate new best practices (e.g., overly thick skill files that should be routers) and suggests optimizations.

Example Optimization: Last 30 Days Skill

A 2,090-line research skill flagged as too thick. Solution: refactor it into a router that points to specialized research sub-skills, reducing token overhead and aligning with progressive disclosure principles.

Notable quotes

They actually removed over 80% of Claude Code's system prompt with no measurable loss. — Jay E (summarizing Tariq's findings)
It's actually better to just let your AI agent use a bit of judgment in order for you to get better results. — Jay E
Your Claude.md becomes more powerful if you make it function as a router to your tree of files. — Jay E

Action items

  • Update Claude Code to the latest version to access /doctor and new features.
  • Audit your Claude.md: convert it from a monolithic reference into a router that points to department-specific files (content, product, community, etc.).
  • Create sub-routers (e.g., content.md) that guide Claude to relevant skills and files for specific intents.
  • Refactor overly thick skills (>1,000 lines) into routers with specialized sub-skills.
  • Build a design interface system (HTML brand book) and link it to a skill file (e.g., /robo) instead of providing specific design examples.
  • Replace rigid rules in skill descriptions with contextual guidance that allows Claude to use judgment.
  • Convert visual specifications from markdown to HTML artifacts (infographics, brand books) for better clarity.
  • Run /doctor-plus monthly to check your setup against the six new rules and optimize accordingly.
  • Set up a /calibrate routine at the end of productive sessions to auto-update skills, memory, and workflows.

More like this