A2UI Agent Files & Skills Maintenance Skill

Guidelines for keeping agent instruction files (AGENTS.md) and specialized skills (.agents/skills/) synchronized with codebase developments.

Core Principles

Maintain Strict Agent Agnosticism: Write all agent-facing configuration files, baseline instructions, and skill guides in a generic, vendor-agnostic manner. Do not use environment-specific or proprietary names (e.g. "Gemini", "Gemini CLI", codebase_investigator, invoke_agent). Use generic alternatives (e.g. "helper subagent", "context retrieval tool", etc.).
Counteract Recency Bias with Subagents: Before modifying documentation or skills after completing a task, invoke a specialized helper or research subagent to perform an objective codebase audit. Do not rely solely on your recent task memory.
Prioritize Specifications over Stale Documentation: Base all updates on specifications found in specification/ rather than general workspace markdown files. Look at JSON schemas first, catalogs second, and protocol guides third.
Enforce Separation of Concerns: Point agents to primary sources of truth inside AGENTS.md. Maintain brief, step-by-step task recipes in skills; do not duplicate specifications or schema details in skill definitions.
Enforce Monorepo Script Uniformity: Guarantee that all Node.js projects are registered as Yarn workspaces in root package.json and implement canonical targets (build, lint, lint:fix, test, format, format:check).