AI Programming Architect

Problem

AI-assisted coding often produces code before the system is clearly understood: interface fields mismatch, external objects are not validated, static configuration is wrong, unit tests pass, but the app fails at runtime. AI 编程架构师 targets this failure mode by treating ARCHITECTURE.md as the single source of truth: clarify requirements, write and persist the architecture document, implement module by module, and sync code changes back into the document.

How It Works

The skill advances through five gated stages:

• S0 Requirement clarification: ask about goals, scale, constraints, and other key gaps before writing anything.
• S1 Architecture selection: compare 2–3 project-specific options and avoid blind copying or over-design.
• S2 Document landing: produce ARCHITECTURE.md and key ADR entries, then write the file into the project root.
• S3 Task decomposition: split the architecture into implementable modules and an ordered task list.
• S4 Module implementation: implement one module at a time, run self-checks and smoke tests, then update code and documentation.

During implementation, it emphasizes exact interface contracts, defensive checks for external objects, integration-path validation, static-data validation on load, and smoke verification of core paths to reduce silent bugs that pass compilation or linting.

Boundaries

It fits tasks that need architectural constraints, documented decisions, and confirmation checkpoints, including existing projects that first need a baseline ARCHITECTURE.md. It is not a shortcut for “skip all discussion and generate a full project”; even when users want fast output, it still requires minimal clarification and document persistence.