← ClaudeAtlas

orchestrator-modelisted

Use when the user says `/orchestrator-mode`, asks you to coordinate agents, or gives large multi-phase work needing parallel delegation, reviews, PRs, scratchpads, or handoffs. Inputs - project goal, repo context, Solo agents/processes, branch/worktree constraints, locked user decisions. Do not use for a direct small edit, a normal code review, or single-agent implementation; use focused coding/review skills instead. Produces delegation plan, agent briefs, monitoring protocol, review/merge routing, state hygiene rules. Escalate on unclear scope, product direction, merge authority, destructive cleanup, or tool availability. Delegates are Solo processes: prefer Solo MCP tools, falling back to the `solo` CLI only when MCP is unavailable - never the in-process Task/Agent tool.
Naoray/skills · ★ 1 · AI & Automation · score 62
Install: claude install-skill Naoray/skills
# Orchestrator Mode **Evidence tier**: P **Basis**: Practitioner-backed multi-agent software coordination, code review gating, worktree isolation, and durable state hygiene. **Source IDs**: Solo CLI + MCP workflow conventions; Anvil worktree workflow; Naoray/skills orchestrator-mode prior art. **Reviewed**: 2026-05-15 You are the coordinator. Your primary output is delegation — not file edits. ## Transport This skill describes the orchestrator role. Tool-specific syntax (spawn, push, durable state) lives in a transport guideline (e.g. [references/transports/solo/README.md](references/transports/solo/README.md) for Solo). The skill is transport-agnostic — the same coordination flow works for any agent runner that supports spawn + push-based reporting. ## TL;DR — operating summary 1. **Read, don't write.** Coding/refactor/test work goes to a delegate. You scope, brief, monitor, evaluate. 2. **Agent selection**: **Codex** for non-visual coding & code-review gates, **Claude** for slash-commands, spec-heavy work, AND all visual/frontend implementation + visual verification (taste/look judgment code review can't give), **Gemini** for adversarial second-opinions and high-stakes plan reviews. 3. **Parallel coding delegates: one working tree each.** Isolation is the rule; the mechanism is your choice (native `git worktree`, `anvil-agent` skill, or Claude's built-in `isolation:'worktree'`). Never share a working tree between parallel coding agents. 4. **Feedback lives in durable