Claude-Modes

Purposeful skill loading for Claude Code. Stop wasting context on skills you don't need.

The Problem

SKILL.md files in Claude Code setups suffer from "drift" and "rot":

• Skills accumulate without organization
• Overlapping functionality wastes context tokens
• No clear loading strategy (everything loads = token waste)
• Users don't know what skills they have or when to use them

The Solution

Modes are operating contexts that load only relevant skills on demand.

modes activate code      # Load design-principles, feature-complete
modes activate security  # Load threat-modeling, vulnerability-detection
modes activate research  # Load evidence-evaluation, research-synthesis

Foundation skills (asking-users, context-gate) ALWAYS load. Mode skills stack on top.

Quick Start

# Install
curl -sL https://raw.githubusercontent.com/CodeTonight-SA/claude-modes/main/install.sh | bash

# List available modes
modes list

# Activate a mode
modes activate code

# Check current mode
modes status

# Deactivate
modes deactivate

How It Works

[Mode Skills]        ← Stacks on top (supplements)
       ↓
[Foundation Skills]  ← ALWAYS ACTIVE (never overridden)

Foundation Skills (Always Active)

Skill	Purpose
asking-users	AskUserQuestion protocol for confidence gates
context-gate	Context window monitoring

Mode Skills (On Demand)

Each mode loads domain-specific skills:

Mode	Skills	Token Budget
code	design-principles, feature-complete, code-agentic	~6000
security	threat-modeling, vulnerability-detection, sanitizing-releases	~3500
research	research-synthesis, evidence-evaluation	~3000

Creating Your Own Mode

modes create my-mode

This creates a template at ~/.claude/modes/definitions/my-mode.yaml:

name: my-mode
description: One-line description. Use when {context}.
version: 1.0.0

triggers:
  commands:
    - /mode my-mode
  patterns:
    - "keyword1"
    - "keyword2"

skills:
  always:
    - skill-a
    - skill-b
  optional:
    - skill-c

token_budget: ~3500
foundation: true

prompt_injection: |
  [MODE: MY-MODE]

  Context-specific instructions...

context:
  - Use case 1
  - Use case 2

See docs/CREATING-MODES.md for the complete guide.

Skill Analyzer

Analyze your existing skills and get mode suggestions:

modes analyze

Output:

Analyzing skills in ~/.claude/skills/...

Skill: feature-complete
  Suggested mode: code (confidence: 0.92)

Skill: threat-modeling
  Suggested mode: security (confidence: 0.87)

Skill: research-synthesis
  Suggested mode: research (confidence: 0.95)

Uncategorized skills (consider creating new mode):
  - custom-skill-1
  - custom-skill-2

CLI Commands

Command	Description
modes list	List available modes
modes activate <name>	Activate a mode
modes deactivate	Deactivate current mode
modes status	Show current mode
modes analyze	Analyze skills, suggest modes
modes create <name>	Create new mode from template
modes sync	Sync with upstream

Token Savings

Loading only relevant skills saves 30-35% of your context budget:

Scenario	Tokens Used	Notes
All skills loaded	~15000	Everything, always
Code mode	~6000	Just development skills
Research mode	~3000	Just research skills
Foundation only	~1000	Minimal, always active

Advanced: Semantic Analyzer

For more accurate skill categorization, enable semantic embeddings:

# Install dependencies
pip install sqlite-lembed sentence-transformers

# Enable semantic mode
modes analyze --semantic

See docs/EMBEDDINGS.md for setup details.

Contributing

We want YOUR modes.

Claude-Modes ships with 3 starter modes. The magic happens when the community builds more.

See CONTRIBUTING.md for how to contribute your modes.

License

Apache 2.0 - See LICENSE

Related Projects

• CIPS - Full Claude Instance Preservation System (includes modes + much more)