A multi-agent workflow that invests planning time upfront using specialized agents, producing detailed specifications that enable single-pass implementation.
## Workflow Overview
| Phase | Duration | Agents | Output |
| ------------------- | -------- | -------------------------- | -------------------------- |
| **Human Draft** | 10 min | Human | Bullet point plan |
| **Refinement Loop** | ~1 hour | Opus 4.6 ↔ Codex 5.4 xhigh | 200-500 line detailed spec |
| **Implementation** | Variable | Claude Code OR Codex | Working code |
| **Review** | Variable | The other agent | Review feedback |
| **Cleanup** | Variable | Custom /deslop agent | Clean code |
| **Debug** | Variable | Any agent | Working deployment |
## Key Principles
1. **Specialized agent roles** — Different models/agents for planning vs. implementation vs. review
2. **Plan amplification** — 6x time investment in planning (10 min → 60 min) produces 10-25x spec detail
3. **Cross-verification** — Implementation agent ≠ review agent provides independent verification
## Implementation Notes
- Markdown serves as the shared interface between human and agents
- The "deslop" step suggests custom post-processing for style/quality enforcement
- This pattern works because the refined spec removes ambiguity that would cause implementation iteration
## Related Concepts
- [[Agent Architecture as Cost Optimization Lever]] — similar boundary discipline
- [[Multi-Agent System Design Patterns]] — role specialization
- [[Specification-Driven Development]] — formal methods parallel
---
*Extracted from: [[My coding workflow is currently -.md]]*