A directory standard for defining intelligent team members.
team/content-lead/
MEMBER.md # role, persona, permissions
skills/ # planner, writer, copy-editor
tools/ # x-search API
context/ # domain knowledge
journal/ # what it learned over time
Team members start from scratch every session. No memory of what worked, no sense of what you prefer, no way to build on past runs. Meanwhile, standards exist for the pieces (SKILL.md, MCP, AGENTS.md) but not for the complete worker.
MEMBER.md bundles persona, skills, tools, context, and a journal into one portable directory. The spec is operator-agnostic -- the same member definition works whether operated by a human, an AI, or both.
journal/ tracks what was done and what the operator preferred. The member reads past entries before starting. This is what makes a member stateful.
skills/ contains SKILL.md files for the capabilities this member owns.
tools/ documents the APIs, CLIs, and integrations this member can use.
context/ holds domain knowledge: playbooks, SOPs, reference docs.
See SPEC.md for the full specification.
---
name: sdr
role: Sales Development Rep
---Two required fields. Everything else is optional. Add skills/, tools/, context/, and journal/ when you need them. A flat file (team/sdr.md) works for simple members.
Use the template to get started.
Start with content-lead/, which has skills, tools, and a journal with real entries. Then look at sdr/ for a full member with context docs.
| Member | Format | Description |
|---|---|---|
| content-lead/ | directory | Writing pipeline with journal that tracks voice preferences |
| sdr/ | directory | Sales member with skills, tools, and context |
| product-manager/ | directory | PM with roadmap context and stakeholder workflows |
| chief-of-staff/ | directory | Stateful briefing agent with journal loop |
| navigator/ | directory | Planner with local skills |
| executor.md | flat file | Builder with abstract skill references |
| validator.md | flat file | Reviewer with approval gates |
MEMBER.md is markdown. Point your AI tool at it.
You are the content-lead. Read team/content-lead/MEMBER.md before doing anything.
For multi-agent setups, each subagent reads a different member:
├── Task: "Act as navigator, read team/navigator/MEMBER.md, plan this feature"
├── Task: "Act as executor, read team/executor.md, build from the plan"
└── Task: "Act as validator, read team/validator.md, review the changes"
Works with Claude Code (CLAUDE.md), Codex (AGENTS.md), Cursor (.cursorrules), OpenClaw, or anything that reads markdown.
MIT