V3 Admin Vite Adds Project Conventions and Skills to Stop AI Coding Drift
AI coding tools still produce unreliable output on real projects because they lack project-specific context. Baking conventions and task workflows directly into a repo gives teams a repeatable way to get consistent AI-generated code across different editors, without repeating instructions in every prompt.
V3 Admin Vite, an open-source Vue3 admin template, has added a layer of AI harness engineering to improve code quality from AI coding tools. The update introduces project-level convention files (AGENTS.md and CLAUDE.md) that define commands, architecture, and coding rules, kept under 200 lines to stay focused. These files are linked via symlinks so a single source of truth works across Claude Code, Cursor, Trae, and other tools.
A set of five project-specific Skills now covers the most repetitive admin-system tasks: creating CRUD pages, updating global stores, modifying routing tables, and using built-in utilities and composables. Each Skill is capped at 500 lines and stored in a shared `.agents/skills/` directory, again symlinked to tool-specific locations. Community Skills from Anthony Fu's collection — covering Pinia, Vite, Vue Router best practices, and testing — are also integrated and installable on demand via a lockfile and a single npx command.
The result is a template that gives AI agents explicit, scoped instructions rather than leaving them to guess project conventions. Early practice shows this layered approach — always-on conventions, file-scoped rules, and on-demand skills — produces more reliable AI-generated code across multiple editors.
Treating AI coding instructions as project source code — versioned, symlinked, and scoped — is a practical shift from ad-hoc prompting to engineered reliability.
The distinction between always-on conventions, file-scoped rules, and on-demand skills mirrors how human developers actually consume project documentation, which likely explains why this structure works across multiple AI tools.
Abstracting repetitive admin-system tasks into Skills turns AI agents from general-purpose coders into domain-aware tools that respect existing patterns instead of inventing new ones.