This is a feature request for Anthropic's Claude Code to support `AGENTS.md` and `.agents/skills/` for defining agent behavior and skills. The community has repeatedly asked for this feature since August 2025, with one issue accumulating over 3,000 upvotes and 200 comments, yet no official response from Anthropic.
A starter prompt for Claude Code, what you'll need, and how to reach them.
You are a senior software engineer. The `anthropics/claude-code` repository is a proprietary codebase, but for this exercise, assume I have cloned the codebase and am able to submit a pull request. Your task is to implement support for `AGENTS.md` and `.agents/skills/` as requested in issue #31005 (which references #6235). Goal: 1. **Implement `AGENTS.md` parsing**: Create logic to parse an `AGENTS.md` file (similar to a `README.md` but for agents) in the root of a project. This file should allow developers to define multiple agents using a simple markdown-based syntax (e.g., `# Agent Name Goal: ... Constraints: ... Skills: skill_name1, skill_name2`). Focus on a clear, human-readable format. 2. **Load `.agents/skills/`**: Implement functionality to discover and load Python or JavaScript skill files (e.g., `my_tool.py`, `another_skill.js`) from a `.agents/skills/` directory within the project. Each skill file should expose a callable function or class that the agent can invoke. 3. **Integrate with agent execution**: Modify Claude Code's agent runtime to use these definitions. When an agent is invoked, it should load its configuration from `AGENTS.md` and have access to the skills defined in `.agents/skills/`. **Stack**: Assume a Python/TypeScript mono-repo structure typical for AI development tools. Prioritize Python for agent logic and skill loading. Focus on minimal dependencies and clean integration. **MVP**: For the first iteration, focus on: - Parsing a single agent definition from `AGENTS.md`. - Loading a single Python skill from `.agents/skills/`. - Allowing this single agent to *recognize* and *attempt to call* the loaded skill, even if the actual tool execution is a mock. **Build/Verify Gate**: The agent (as defined in `AGENTS.md`) should log that it has successfully parsed its definition and discovered the skill from `.agents/skills/`. A simple test agent that attempts to 'use' a mock skill should demonstrate the integration. Provide the necessary code snippets for `AGENTS.md`, a mock skill file, and the modified Claude Code internal logic to demonstrate the end-to-end flow.
Reach out to the community members discussing `AGENTS.md` support on the Claude Code GitHub issue, highlighting how this work directly enables more robust agent development and testing using `agent-eval-lab` and `mcp-kit`.
# Support for AGENTS.md and .agents/skills/, the community has been asking since August 2025 ## Preflight Checklist - [x] I have searched existing requests and this feature has been requested **multiple times** with **zero official response** - [x] This is a single feature request ## Problem Statement This has been raised repeatedly by the community for **over 7 months** with no acknowledgment from the Anthropic team: - #5659, Configurable Agent Definition File Path (Aug 2025, closed by bot) - #6235, Support AGENTS.md (Aug 2025, **3,020 upvotes, 224 comments, still open, zero team response**) - #14474, Will Claude Code support AGENTS.md in the future? (Dec 2025, auto-closed as duplicate) - #20820, Reserve skills/ for user-managed skills (Jan 2026, auto-closed for inactivity) - #22022, Claude Code ignores AGENTS.md rules (Jan 2026, auto-closed as duplicate) - #25882, Auto-load ~/.claude/AGENTS.md (Feb 2026, marked stale) **Not a single one has received a response from anyone at Anthropic.** The main issue (#6235) has over three thousand upvotes and two hundred comments from the community, and the only responses are from a bot closing duplicates. 7 months. 3,020 upvotes. Zero acknowl
Standard for any GitHub contribution.
Core language for Claude Code and AI tool development.
Required to understand integration points and existing agent runtime – ~2-3 days of code spelunking.
Learn it: Anthropic docs ↗
Get set up: Create a key at console.anthropic.com → set `ANTHROPIC_API_KEY` → `npm i @ai-sdk/anthropic`.
Comment on the GitHub issue #31005 (or #6235) and, once a working prototype is available, open a draft Pull Request to demonstrate the solution.
“I've seen the strong community demand for `AGENTS.md` and `.agents/skills/` support and have built a working prototype for how this could be implemented within Claude Code. Here's a link to my fork/demo – happy to discuss and potentially open a PR.”
Open the original ↗