README

Claude Code subagent imported from leogodin217/leos_claude_starter (.claude/agents/README.md). Copyright stays with the author.

Agents and Skills

Agents are workers with specific expertise, tools, and constraints, invoked via the Task tool. Skills are processes that orchestrate work, invoked via the Skill tool (slash commands). Skills dispatch agents.

Shared agents (symlinked into adopting repos)

These bodies are repo-neutral by contract:

• No repo principle content. Repo principles reach agents two ways: the SubagentStart hook injects the repo's worker-protocol.md (which carries the repo's config-boundary section), and agents read the repo's CLAUDE.md § Core Principles at task start.
• No principle citations by number. Numbering is repo-local and collides across repos; shared bodies cite principles by name only.
• No repo vocabulary. Prose names roles (author, source tree, output judge), never one repo's instances.

Each adopting repo adds its own output-judge agent (named by the output-judge-agent shape key) — an agent that reads program output rather than source, listed in the inject hook's CODE_NAV_EXEMPT.

See ADOPTION.md at the repo root for the full contract.

Adding an agent

Create .claude/agents/<name>.md:

---
name: agent-name
description: Brief description for the Task tool
tools: Tool1, Tool2, Tool3
model: sonnet
---

You are the {Role}. {One-sentence job.}

## What You Produce
## What You Do NOT Do

Design rules: minimal context, explicit "do not" constraints, exact output format, only the tools the agent needs. If the agent is general — useful to any adopting repo — it belongs here and must meet the repo-neutral contract above. If a sentence in it cannot be written without naming one repo's nouns, it is a repo agent; keep it in that repo.