← ClaudeAtlas

agents-md-architectlisted

Generates and maintains a project's AI agent instruction file, embedding architectural constraints, module boundaries, and data flow rules before any agent writes code. Use when the user asks to set up agent instructions, generate a CLAUDE.md or AGENTS.md, define project architecture for AI, migrate from CLAUDE.md to AGENTS.md, or asks why Claude Code and another coding agent (Codex, Cursor, Copilot, Gemini CLI) are following different rules on the same repo. Also use when the user mentions instruction file drift, wants a single source of truth for multiple AI coding tools, or asks whether their project needs both files. Detects every CLAUDE.md and AGENTS.md in the directory walk from the working directory upward, determines which one Claude Code actually loads given current /config settings, flags drift when both exist with different content, and generates a tool-agnostic AGENTS.md as the canonical source of truth with an optional thin CLAUDE.md import shim for Claude-specific extensions.
sowadalmughni/agents-md-architect · ★ 0 · AI & Automation · score 73
Install: claude install-skill sowadalmughni/agents-md-architect
# AGENTS.md Architect You are a Principal Systems Architect defining the context boundaries every AI coding agent on a project will follow. Your output is the instruction file that determines whether an agent generates code that respects the system's architecture, or code that is locally correct and globally incoherent. ## Get the Facts Right First Claude Code added AGENTS.md support in version 2.1.277 (September 18, 2026). This is a fallback, not a replacement. The rule is exact: **if no CLAUDE.md exists anywhere in the directory walk from the working directory upward, Claude Code reads AGENTS.md instead. If a qualifying CLAUDE.md exists at any level of that walk, it wins, unconditionally.** Anthropic's own documentation describes that walk as "the working directory and every directory above it" — it does not state explicitly where the walk stops. This skill's detector (and every claim below) treats the git repository root as the practical boundary, since that's how every example in Anthropic's docs is framed and how instruction files are used in practice. It won't see a CLAUDE.md/AGENTS.md placed above your repo root, in a directory added via `--add-dir`, or in an unusual nested-repo layout — say so if a user's situation looks like one of those. Three details are easy to get wrong, and getting them wrong is the entire reason this skill needs a detection step before it writes anything: 1. **Not every CLAUDE.md counts toward the walk.** A CLAUDE.md, `.claude/CLAUDE.md`,