member-runtime_skill

Overview

This skill loads and runs MEMBER.md team members—complete AI workers that bundle a persona, skills, tools, context, and a growing journal. It activates a member, enforces their permissions and workflow, and records stateful journal entries so the worker improves over time.

How this skill works

When you activate a member (for example, "be my chief of staff"), the runtime locates the MEMBER.md definition, parses its frontmatter for name, role, skills, tools, and permissions, then loads local skills and context files. It reads the member's journal and memory to adopt current preferences, enforces declared permissions during execution, and appends a dated journal entry after tasks to capture outcomes and emerging patterns.

When to use it

• Activate a persistent persona to perform recurring tasks with consistent rules and style
• Delegate multi-step workflows that require tool access, role-specific permissions, or a memory of past interactions
• Run a stateful AI worker that learns user preferences over time
• Switch between distinct team members while preserving separate journals and settings
• Draft or execute actions that must respect explicit approval or send restrictions

Best practices

• Define clear frontmatter: name, role, skills, tools, and explicit permissions to avoid ambiguity
• Include domain context files (playbooks, preferences) so the member has immediate reference knowledge
• Use the journal each run—record tasks, deliverables, and observed user reactions to build durable preferences
• Treat permissions as authoritative: draft instead of sending for any can-*: false flags and request approval when required
• Promote patterns to durable memory only after consistent evidence (three or more days) to avoid premature assumptions

Example use cases

• Start a chief-of-staff member that triages inbox items, drafts responses, and schedules follow-ups with approval rules
• Activate an SDR persona that uses sales playbooks from context files and tracks outreach patterns in the journal
• Run a product-research member that loads domain docs, queries tools, and records findings to build a searchable memory
• Switch between members for different projects to keep personas, skills, and learned preferences isolated
• Install a community manager who drafts posts but requires approval before publishing due to permission constraints