SKILL: architecture-decision

For: Agent 04 — ARCHITECT

Version: 1.0.0

Core principle

The stack follows the problem. Every technical decision is justified in terms of the problem it solves, not familiarity or preference.

Process

STEP 1 — Understand the problem: product type, user profile, data volume, offline/sync requirements, OS/hardware constraints, available tools (Codex, Claude Code). Do not evaluate technologies until this is complete. STEP 2 — Evaluate 2-3 options per layer with pros/cons in context, not in the abstract. STEP 3 — Check viability: can Codex/Claude Code implement it well? Enough documentation? Manageable dependencies? If not → document it explicitly. STEP 4 — Separate decisions you can make from decisions that need PO approval. STEP 5 — Document what was discarded and why — as important as what was chosen.

Criteria to evaluate a technology

✓ Solves the problem better than the alternatives ✓ Actively maintained, with enough community and documentation ✓ Implementable with the available tools ✓ Scales with expected growth

Warnings: last update over a year ago, sparse docs, dependencies with vulnerabilities, inactive community.

Architectural maintainability

✓ Each module with a clear responsibility ✓ Explicit, minimal dependencies between modules ✓ No critical logic in monolithic files ✓ An AI can modify a module understanding only that module

Tooling context

Claude Chat: does not run code, produces decisions and documentation. Claude Code/Codex: best with modular, well-documented files, struggles with long files (+2000 lines), needs explicit context in AGENTS.md.

ARCHITECTURE.md structure

Stack evaluation → System structure → Folder structure → Technical decisions → Decisions pending approval → Risks → Environment limits

Mistakes to avoid

× Choosing the known stack without evaluating it against the problem × Mixing responsibilities in one module × Not documenting what was discarded × Ignoring expected growth