Skill Agent Architect

Problem

In Claude Code and Codex ecosystems, skill packages often become too broad, overlap with existing agents, or consume context budgets that make routing brittle. Editing prompts alone usually misses the structural issue: agent failures are often context-architecture failures.

How It Works

Architect treats skill creation and refinement as an engineering workflow:
- CREATE: follows UNDERSTAND → ENVISION → ANALYZE → DESIGN → GENERATE → VALIDATE to produce a complete package with SKILL.md, reference files, capability summaries, collaboration contracts, and explicit input/output boundaries.
- IMPROVE / COMPRESS: scores health and overlap first, then compresses context while preserving behavioral, structural, integration, and routing equivalence.
- EVOLVE: handles self-improvement only through INTROSPECT → DIAGNOSE → PRESCRIBE → MUTATE → VERIFY → PERSIST, with rollback snapshots and budget approvals.

It front-loads capability fit, collaboration boundaries, and topology choices, rather than blaming failures on prompt wording. For multi-agent designs, it checks for formal topology before allowing loosely coupled agent stacks.

Boundaries

This skill is not a replacement for task-chain orchestration, product delivery, or application architecture analysis. If overlap exceeds 30%, compression exceeds 20%, routing changes materially, or Level C self-modification is involved, confirmation is required. Validation, health scoring, and equivalence checks must not be skipped.