AI-Assisted Development

This template distills "how an AI assistant should work within this architecture" into executable rules and skills, stored in the .trae/ directory and evolving alongside the code.

Composition

Item	Path	Responsibility
Architecture development rules	.trae/rules/architecture.md	The machine-enforced constraint layer of the architecture: layered call chain and prohibitions, plugin inventory (single source of truth for import paths), standard implementation order, type organization criteria, UTS red lines, unique assembly points, verification DoD
AI behavior protocol	.trae/rules/core-protocol.md	Critical execution, work modes (interactive / unattended / plan-first), DoD checklist, lessons-learned records
Document-driven workflow	.trae/rules/workflow.md	Mandatory feature implementation flow: create task document → wait for and read answers → implement and verify → backfill implementation records, all persisted to disk across sessions
General convention distillation	.trae/rules/project_rules.md	Excerpts of uni-app x / UTS / uvue / ucss general constraints
New feature skill	.trae/skills/arch-feature/SKILL.md	Six-step implementation: requirement confirmation → DTO → api → store → pages/components → assembly point check → verification loop
Task document archive	.trae/documents/	Two-level archive by "requirement type / date"; implementation-log.md is the project-wide task index
MCP config	.trae/mcp.json	uni-app-x MCP server for querying plugin and API platform compatibility

Design Principles

• Principle layer and enforcement layer separated: the repository homepage architecture document explains "why it is designed this way", while .trae/rules/architecture.md specifies "how the AI must write" — the two share the same origin and evolve in sync;
• Red lines never drift: AI-generated code is constrained within the layering; the UTS red-line checklist is injected directly into context, catching compile-level pitfalls at generation time;
• Document-driven: requirement analysis, pseudocode, questions and answers are all persisted to task documents, iterating across sessions; conversation only handles notification and execution.

Practical Benefit

The biggest benefit is not "writing faster" but architecture conventions enforced by machines: the AI does not bypass layering for "brevity", does not write plugin import paths from faulty memory, does not fall into the pitfalls catalogued in the UTS red-line checklist — precisely the parts of human-machine collaboration that are easiest to lose control of.