Agent Skills › BuilderIO/skills › efficient-orchestrator

efficient-orchestrator

GitHub

一种多模型协作模式，主模型负责规划与审核，廉价子代理处理研究、编码和测试等具体任务，旨在优化成本并提升效率。

skills/efficient-orchestrator/SKILL.md BuilderIO/skills

Trigger Scenarios

用户显式调用 /efficient-orchestrator 或其别名

Install

npx skills add BuilderIO/skills --skill efficient-orchestrator -g -y

More Options

Use without installing

npx skills use BuilderIO/skills@efficient-orchestrator

指定 Agent (Claude Code)

npx skills add BuilderIO/skills --skill efficient-orchestrator -a claude-code -g -y

安装 repo 全部 skill

npx skills add BuilderIO/skills --all -g -y

预览 repo 内 skill

npx skills add BuilderIO/skills --list

SKILL.md

Frontmatter

{
    "name": "efficient-orchestrator",
    "description": "Run only when the user explicitly invokes \/efficient-orchestrator (or the \/efficient-frontier and \/efficient-fable aliases). Keeps an expensive model (Fable, Opus, Astra, Sol) on planning, judgment and review while cheaper subagents do bounded research, coding, QA and log work. Never self-activate.",
    "disable-model-invocation": true
}

Efficient Orchestrator

Use this only when the user asks for it by name. Without that request, do the work yourself and don't spawn helpers.

The expensive main model plans, decides, integrates and reviews. Cheaper subagents do the token-heavy, bounded work.

Model tiers

Main model	Cheap helpers	Use
Claude Fable / Opus	`haiku`	search, scans, docs extraction, log reduction, mechanical edits, test and browser runs
	`sonnet`	coding or analysis that needs judgment
OpenAI Astra / Sol	`gpt-6.1-terra` low or medium	coding, QA, research, most delegated work
	`gpt-6-luna`	bulk mechanical scans, only when Terra isn't needed

Set the model and effort explicitly on every spawn; never let a helper inherit the expensive model. If the main model is already a cheap tier, don't use this skill.

Workflow

Keep at the top: architecture, prioritization, ambiguity, risk, synthesis and final review.
Delegate: repo scans, research, narrow patches with clear file ownership, test and browser passes, failure clustering, log reduction.
Write each handoff as a self-contained packet: repo path, objective, scope and out-of-scope, files or search targets, verification command, return format, stop conditions.
Require compact returns: findings, changed files, commands run, evidence, open decisions.
Inspect the evidence that matters before claiming done. Delegated output is evidence, not a verdict.

Limits

At most 3–5 helpers at once. Helpers never spawn their own helpers.
Close each helper as soon as its result is in.
Don't delegate the blocker your next step depends on, and never give two helpers the same files.
Stop helpers that return nothing usable twice; do that step yourself.
When done, say roughly how much work was delegated, so the cost is visible.