Adaptive Claude Agents - Skill Instructions

Overview

Skill Name: Adaptive Claude Agents Version: 1.1.0 Purpose: Auto-generate project-specific Claude Code subagents that adapt to your development phase and framework.

This skill enables Claude to:

1. Automatically detect project technology stack (11 frameworks supported)
2. Detect development phase (Prototype, MVP, Production)
3. Generate appropriate subagents with AGGRESSIVE delegation policy
4. Provide framework-specific workflow guides

When to Use This Skill

Use this skill when:

• User asks: "Analyze my project and generate subagents"
• User asks: "What development phase am I in?"
• User mentions: "Set up adaptive agents for this project"
• Starting work on a new project that needs subagent configuration

DO NOT use for:

• Simple file operations
• General code review
• Tasks unrelated to project setup

Available Commands

1. Project Analysis & Subagent Generation

Command:

python3 "$SKILLS_DIR/adaptive-claude-agents/skills/project-analyzer/analyze_project.py" [PROJECT_PATH] [--auto]

When to use:

• First time setting up a project
• After updating Adaptive Claude Agents to latest version
• User explicitly requests project analysis

What it does:

1. Detects tech stack (Next.js, FastAPI, Go, Flutter, etc.)
2. Generates .claude/agents/SUBAGENT_GUIDE.md with framework-specific workflows
3. Creates specialized subagent templates
4. Adds AGGRESSIVE policy to project's CLAUDE.md

Output:

✅ Successfully generated subagents!
📁 Location: .claude/agents/
📖 Usage Guide: .claude/agents/SUBAGENT_GUIDE.md

2. Development Phase Detection

Command:

python3 "$SKILLS_DIR/adaptive-claude-agents/skills/adaptive-review/detect_phase.py" [PROJECT_PATH]

When to use:

• User asks about current development phase
• Determining appropriate code review rigor
• Project planning discussions

What it does:

1. Analyzes version number, commits, tests
2. Classifies as Prototype (3/10), MVP (6/10), or Production (10/10)
3. Provides phase-appropriate review recommendations

Output:

Phase: MVP
Confidence: 72%
Review Rigor: 6/10 (moderate)

3. Framework Detection Only

Command:

python3 "$SKILLS_DIR/adaptive-claude-agents/skills/project-analyzer/detect_stack.py" [PROJECT_PATH]

When to use:

• Quick framework check without generating subagents
• Debugging detection logic
• User asks: "What framework does this project use?"

What it does:

1. Detects framework with confidence score
2. Lists detected tools (testing, database, etc.)
3. Shows detection reasoning

Output:

Detected: Next.js 14
Confidence: 98%
Tools: Vitest, Testing Library, Tailwind CSS

4. Update Existing Subagents

Command:

python3 "$SKILLS_DIR/adaptive-claude-agents/skills/project-analyzer/analyze_project.py" . --[MODE] --auto

When to use:

• User asks: "サブエージェントを更新して" / "Update my subagents"
• User mentions: "最新のテンプレートに更新" / "Update to latest templates"
• After framework update (e.g., bash update.sh)
• User asks: "新しいテンプレートがあるか確認して" / "Check for new templates"

Available Modes:

1. --update-only (RECOMMENDED - Safest)

   • Updates only existing .claude/agents/*.md files
   • Preserves file list (no new files added)
   • ⚠️ Overwrites customizations
   • No backup created
   • Use when: User hasn't customized agents

2. --merge (SAFE - Preserves customizations)

   • Adds new templates from latest version
   • Preserves existing files (no changes)
   • ✅ Customizations protected
   • Creates backup: .claude/agents.backup.YYYYMMDD-HHMMSS/
   • Use when: User has customized agents and wants new templates

3. --force (DESTRUCTIVE - Complete regeneration)

   • Regenerates all files from templates
   • ❌ Customizations lost
   • Creates backup
   • Use when: User explicitly requests complete refresh

Interactive Workflow:

When user asks to update subagents:

1. Check for existing agents:

   bashCopy

   ls .claude/agents/*.md 2>/dev/null
   

2. Ask user which mode:

   Copy

   I found existing subagents in .claude/agents/. Which update mode would you like?
   
   1. --update-only (Recommended): Update existing files to latest versions
      ⚠️  Your customizations will be overwritten
      ✅ Safest for non-customized projects
   
   2. --merge: Add new templates, preserve your existing customizations
      ✅ Your changes are safe
      ✅ Gets new templates
      ⚠️  Existing files won't get template improvements
   
   3. --force: Complete regeneration (creates backup first)
      ❌ Customizations lost
      ✅ Fresh start with latest templates
   

3. Execute with chosen mode:

   bashCopy

   python3 "$SKILLS_DIR/adaptive-claude-agents/skills/project-analyzer/analyze_project.py" . --[MODE] --auto
   

4. Report results:

   Copy

   📊 Update Summary:
   • Updated: 3 existing agent(s)
   • Preserved: 1 customized agent(s)
   • Backup: .claude/agents.backup.20251024-073851/
   

Default recommendation: If user doesn't specify, suggest --merge as the safest option.

Slash command shortcut: User can also use /update-subagents

Supported Frameworks

Total: 11 frameworks, 15 specialized templates

AGGRESSIVE Mode Philosophy

Core Principle: Installation = Consent to Proactive Subagent Usage

Mandatory Delegation Rules

Claude MUST use Task tool when:

1. 3+ files need similar modifications

   • Example: "Apply blur fix to assessment.js, soap.js, nursing_plan.js"
   • Subagent: general-purpose
   • Time saved: 30-60 minutes

2. Searching entire codebase for patterns

   • Example: "Find all uses of version_number"
   • Subagent: Explore (thoroughness: "very thorough")
   • Time saved: 60-90 minutes

3. E2E testing or automated verification

   • Example: "Test login flow → API call → DB validation"
   • Subagent: general-purpose + chrome-devtools-tester
   • Time saved: 45+ minutes

4. 2+ independent tasks can run in parallel

   • Example: "Update .gitignore + Refactor UI components"
   • Subagent: Multiple general-purpose in single message
   • Time saved: 30+ minutes

Cost vs Time Guidance

Rule of Thumb: 20k token overhead is acceptable for 30+ minutes saved.

Input/Output Specifications

analyze_project.py

Inputs:

• PROJECT_PATH (optional, default: current directory)
• --auto flag (skip user confirmation)
• --verbose flag (enable debug logging)

Outputs:

• .claude/agents/SUBAGENT_GUIDE.md (~250 lines, framework-specific)
• .claude/agents/*.md (specialized subagent templates)
• Updated CLAUDE.md (AGGRESSIVE policy added)

Exit codes:

• 0: Success
• 1: Detection failed or user aborted

detect_phase.py

Inputs:

• PROJECT_PATH (optional, default: current directory)

Outputs:

• JSON or human-readable phase analysis
• Confidence score (0-100%)
• Review rigor recommendation (1-10)

detect_stack.py

Inputs:

• PROJECT_PATH (required)
• --help (show usage)

Outputs:

• Framework name
• Version (if detected)
• Confidence score
• Detected tools by category

Example Workflows

Workflow 1: New Project Setup

User: "I started a new Next.js project. Set up adaptive agents for it."

Claude:
1. Confirms project path
2. Runs: analyze_project.py /path/to/project --auto
3. Reports success with generated files
4. Guides user to read SUBAGENT_GUIDE.md

Workflow 2: Framework Identification

User: "What framework is this project using?"

Claude:
1. Runs: detect_stack.py .
2. Shows: "Detected: FastAPI (80% confidence)"
3. Lists tools: pytest, SQLAlchemy, uvicorn

Workflow 3: Phase-Appropriate Review

User: "Review this new feature"

Claude:
1. Runs: detect_phase.py .
2. Detects: Prototype (3/10 rigor)
3. Applies light review: "Does it work?" focus only

Workflow 4: Multi-File Modification (AGGRESSIVE)

User: "この3つのファイルに同じ修正を適用して"

Claude (Internal):
1. Detects: 3 files = MANDATORY Task tool usage
2. Delegates to: general-purpose subagent
3. Returns: Unified result after parallel execution

(No user confirmation needed - AGGRESSIVE mode)

Limitations

Framework Coverage

• Supported: 11 frameworks (Next.js, FastAPI, Go, Flutter, etc.)
• Not Supported: Ruby on Rails, ASP.NET, Rust (Actix)
• Workaround: Manual .claude/project.yml creation

Detection Accuracy

• Average: 85% confidence
• Range: 70% (Flask) to 100% (Next.js, Python ML)
• False Negatives: Possible for unconventional project structures

Platform Support

• macOS: ✅ Fully tested
• Linux/WSL: ✅ Supported
• Windows (native): ⚠️ Untested (may require path adjustments)

Error Handling

Common Errors

Error: "Could not detect tech stack" Cause: Project doesn't match supported frameworks Solution:

1. Check supported frameworks list
2. Create manual .claude/project.yml
3. Open issue: https://github.com/SawanoLab/adaptive-claude-agents/issues

Error: "Python dependency 'yaml' not found" Cause: pyyaml not installed Solution: python3 -m pip install --user pyyaml

Error: "Permission denied: .claude/agents/" Cause: Insufficient write permissions Solution: chmod +w .claude/agents/ or run with appropriate permissions

Security Considerations

Safe Practices

• ✅ All processing is local (no network requests)
• ✅ No telemetry or analytics
• ✅ Input validation in detection scripts
• ✅ No code execution from detected files

User Responsibilities

• Review generated SUBAGENT_GUIDE.md before using
• Verify CLAUDE.md policy matches team preferences
• Do not commit sensitive data to templates

Dependencies

Required (Bundled)

• Python 3.9+
• pyyaml library (auto-installed by install.sh)

Optional

• Git (for commit/version detection in phase analysis)
• Claude Code (for subagent execution)

Version History

Current Version: 0.4.3-beta Stability: Beta (production-ready with minor edge cases)

Support & Resources

• Documentation: https://github.com/SawanoLab/adaptive-claude-agents
• Migration Guide: docs/MIGRATION_v0.4.3.md
• Examples: docs/EXAMPLES.md
• Troubleshooting: docs/TROUBLESHOOTING.md
• Issues: https://github.com/SawanoLab/adaptive-claude-agents/issues

Notes for Claude

• Be proactive: Suggest running analyze_project.py for new projects
• AGGRESSIVE mode: Use Task tool for 3+ files automatically
• Framework-aware: Read SUBAGENT_GUIDE.md for project-specific workflows
• Progressive disclosure: Only load this skill when project setup is discussed
• User confirmation: Skip with --auto flag, ask otherwise

Remember: Installation = User consent to aggressive subagent usage. Trust the automation!