A Claude Code skill that enables systematic code generation through extreme decomposition, voting, and restartable batch execution.
MAKER (Massively Decomposed Agentic Processes) is a framework that transforms how Claude writes code:
- Decompose - Break tasks into minimal, atomic subtasks
- Generate - Create 3 different solutions for each subtask
- Vote - Evaluate and select the best solution
- Assemble - Combine winning solutions into complete code
- Persist - Save progress for restartability
| Traditional Approach | MAKER Approach |
|---|---|
| One-shot code generation | Systematic decomposition |
| Single implementation | 3 solutions per step, best selected |
| Lost progress on interruption | Checkpointed, restartable batches |
| Variable quality | Consistent, evaluated quality |
Copy the files to your Claude configuration:
# Create directories if they don't exist
mkdir -p ~/.claude/skills/maker-coder
mkdir -p ~/.claude/commands
# Copy skill and command
cp skills/maker-coder/SKILL.md ~/.claude/skills/maker-coder/
cp commands/maker.md ~/.claude/commands//maker Build a REST API with authentication
Use the MAKER framework to build a REST API with authentication
The skill auto-activates for:
- Complex coding tasks (>5 steps)
- Multi-file implementations
- Data analysis on large datasets (>1000 rows)
- ETL pipelines and data transformations
- Production/critical code
User: /maker Create a CSV email validator with export
Claude: I'll use MAKER with batch execution for restartability.
📁 Created task folder: .maker-tasks/csv-validator-20241225/
**Decomposition:** 7 steps in 3 batches
Batch 1 (steps 1-3): Core reading and validation
Batch 2 (steps 4-5): Filtering and writing
Batch 3 (steps 6-7): Orchestration and CLI
━━━ BATCH 1/3 ━━━
✓ Step 1/7: CSV reader - Selected DictReader approach
✓ Step 2/7: Email validator - Selected RFC regex
✓ Step 3/7: Row validation - Selected functional approach
💾 Checkpoint saved. Task can be safely interrupted.
[continues...]
MAKER creates a .maker-tasks/ folder to track progress:
.maker-tasks/
└── csv-validator-20241225/
├── task-manifest.json # Task metadata
├── decomposition.md # Full breakdown
├── progress.json # Step completion status
├── batch-001/ # Solutions per batch
│ ├── step-01-solution.md
│ └── ...
└── final/
└── complete.py # Assembled code
User: Continue the CSV validator task
Claude: Found existing task 'csv-validator-20241225' at batch 2, step 4.
Resuming from: Data filter function
Remaining: 4 steps across 2 batches
Each solution is scored on:
| Criterion | Weight | Description |
|---|---|---|
| Correctness | 30% | Solves the problem correctly |
| Clarity | 25% | Readable, well-documented |
| Robustness | 25% | Handles edge cases |
| Efficiency | 20% | Reasonable performance |
- "Build a REST API with..."
- "Create a data processing pipeline"
- "Write a web scraper with..."
- "Implement a caching layer"
- Multi-step algorithms
- Production code requirements
- Simple single-function tasks
- Quick fixes and typos
- Explanations or questions
- Tasks with <3 steps
maker-skill/
├── README.md # This file
├── skills/
│ └── maker-coder/
│ └── SKILL.md # Main skill definition
└── commands/
└── maker.md # Slash command definition
The skill uses these defaults:
- Batch size: 3-5 steps per batch (scales with complexity)
- Solutions per step: 3
- Checkpoint frequency: After each batch
MIT License - Feel free to use and modify.
Contributions welcome! Please open an issue or PR.