minimum-viable-agentic（最小可行剂）

Minimum Viable Agentic Layer Skill

Guide teams through creating the essential agentic layer components to start agentic coding.

When to Use

• Starting agentic coding in a new project
• Adding agentic layer to existing codebase
• Understanding bare minimum requirements
• Quick-starting agentic automation

Core Concept

"For your minimum viable agentic layer, you really only need these pieces: AI developer workflow directory, prompts, and plans."

Minimum Viable Structure

project/
├── specs/                    # Plans for agents
│   └── (generated plans go here)
├── .claude/
│   └── commands/            # Agentic prompts
│       ├── chore.md         # Chore planning
│       └── implement.md     # Implementation HOP
└── adws/                    # AI Developer Workflows
    ├── adw_modules/
    │   └── agent.py         # Core execution
    └── adw_chore_implement.py  # Gateway script

Total: 4-5 files to bootstrap.

Implementation Workflow

Step 1: Create Directory Structure

mkdir -p specs
mkdir -p .claude/commands
mkdir -p adws/adw_modules

Step 2: Create Chore Template

.claude/commands/chore.md:

# Chore Planning

Create a detailed plan for this chore task.

## Task
$ARGUMENTS

## Output
Create a spec file at: specs/chore-{adw_id}-{name}.md

Include:
- Task description
- Files to modify
- Step-by-step implementation
- Validation criteria

Step 3: Create Implement HOP

.claude/commands/implement.md:

# Implementation

Implement the plan provided.

## Plan File
$ARGUMENTS

Read the plan file and implement each step.
Report changes with git diff --stat when complete.

Step 4: Create Agent Module

adws/adw_modules/agent.py:

• Claude Code subprocess execution
• Request/response data models
• Output file handling
• Error handling

Step 5: Create Gateway Script

adws/adw_chore_implement.py:

• Accept chore description
• Execute /chore to generate plan
• Execute /implement with plan
• Report results

Validation Checklist

After setup, verify:

• specs/ directory exists
• .claude/commands/chore.md exists
• .claude/commands/implement.md exists
• adws/adw_modules/agent.py exists
• Gateway script runs successfully

Time Investment

Scaling Path

After MVP, add progressively:

1. Week 2: Add /bug and /feature templates
2. Week 3: Add hooks for automation
3. Week 4: Add triggers for scheduling
4. Week 5: Add worktree isolation
5. Week 6+: Add full SDLC workflows

Key Memory References

• @agentic-layer-structure.md - Full structure reference
• @gateway-script-patterns.md - Script patterns
• @template-engineering.md - Template design

Output Format

## Minimum Viable Agentic Layer Setup

**Project:** {name}
**Status:** Ready for Implementation

### Files to Create

1. **specs/** (directory)
   - Will contain generated plans

2. **.claude/commands/chore.md**
   - Chore planning template
   - Generates specs from descriptions

3. **.claude/commands/implement.md**
   - Implementation HOP
   - Executes generated plans

4. **adws/adw_modules/agent.py**
   - Core agent execution
   - Subprocess handling
   - Output management

5. **adws/adw_chore_implement.py**
   - Gateway script
   - Composes chore + implement

### Implementation Order
1. Create directories
2. Add chore.md template
3. Add implement.md template
4. Create agent.py module
5. Create gateway script
6. Test end-to-end

### Estimated Time
5-8 hours to production-ready MVP

Common Mistakes

• Over-engineering MVP (too many templates too soon)
• Skipping agent.py (trying to call claude directly)
• No error handling in gateway scripts
• Not organizing output files consistently

Version History

• v1.0.0 (2025-12-26): Initial release

Last Updated

Date: 2025-12-26 Model: claude-opus-4-5-20251101