How to use the Advisor Orchestrator Worker agent skill

What is an agent skill, and what does this one make the agent do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to handle a certain kind of task. The agent reads the instructions when the skill is relevant to your request.

Once installed, this skill makes the agent follow a fixed loop: frame, plan, review, delegate, verify, synthesize, and final review. Instead of trying to do everything in one pass, the agent asks a stronger advisor model to review the plan, sends self-contained subtasks to cheaper worker models in parallel, checks each result against its own acceptance criteria, and then assembles the deliverable.

• Frame: state the deliverable and 3 to 5 checkable success criteria.
• Plan: split the work into subtasks with wave assignments.
• Plan review: a mandatory advisor consult before any work is dispatched.
• Delegate and verify: workers run in parallel, and each result gets a PASS, FIX, or ESCALATE verdict.
• Synthesize and taste pass: the draft gets a second advisor review before it ships.

Install the skill in Claude (ZIP download or Settings upload)

Download the ZIP for this skill from this page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder so it applies only there.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/advisor-orchestrator-worker/
• Project install: .claude/skills/advisor-orchestrator-worker/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload the skill

What you need installed before it works

The skill calls outside models over the network, so it needs command-line tools and keys in place before the run starts. The agent checks these at the frame step and tells you how to fix any gap.

Workers resolve through the Antigravity CLI (agy) first, then a Gemini API key (GEMINI_API_KEY or GOOGLE_API_KEY). The advisor resolves through the claude CLI first, then ANTHROPIC_API_KEY. The skill also needs bash, jq, and perl, which it uses for timeouts because timeout(1) is missing on stock macOS.

• bash shell
• jq
• Antigravity CLI (agy), or GEMINI_API_KEY or GOOGLE_API_KEY
• claude CLI, or ANTHROPIC_API_KEY
• perl

How to trigger the skill with natural requests

You do not need a special command. The skill's description covers tasks that are too large for one model pass, work that needs parallel research or generation across many subtasks, and requests that name the pattern directly. Phrases such as "too big for one model", "fan this out", or "run an advisor-worker loop" are good triggers.

Be specific about the volume of work. The skill is designed for jobs like researching a dozen competitors at once, so naming the number of items and the output you want helps the agent decide the pattern fits.

• "This is too big for one model. Fan this out: research these 12 competitors and give me a comparison table."
• "Run an advisor-worker loop to draft and check a set of product descriptions for our 30 SKUs."
• "Have a stronger model review the plan while cheap workers execute the research on each company."

Examples of good and poor fits

The skill is built for work that splits cleanly into independent pieces. Each worker sees only its own brief, so the subtasks must carry their own inputs and criteria. Work that needs one coherent pass is a poor fit, and the skill itself says one model pass is enough in that case, with a large overhead otherwise.

Do not send sensitive or confidential data through it. Subtasks and briefs are sent to external model APIs over the network.

• Good fit: researching many competitors in parallel, then merging the findings into one report.
• Poor fit: fixing a typo in one file or answering a simple question.
• Poor fit: anything involving confidential data that should not leave your environment.

Safety notes: scripts and network use

This skill runs scripts and makes network calls. Worker briefs are written to temp files and dispatched from empty temp directories with a minimal environment, which reduces leakage but is not a sandbox. Temp files are cleaned up at the end of a run.

If a role has no working CLI or API key, the agent can offer degraded mode, where it temporarily plays that role itself. Outputs produced that way are labeled [DEGRADED: role], and the isolation the design relies on is weakened.

• Runs shell scripts and uses the network.
• Workers and advisor receive your briefs and plan via external APIs.
• Use only with data you are comfortable sending to those providers.
• Without the CLIs or keys, results lose context isolation and are labeled as degraded.

Tips for getting reliable results

Give the agent a clear deliverable and checkable success criteria at the start, since the frame step depends on them. Set a budget if you want to control cost; the skill suggests a budget shape but it is a guideline rather than a hard limit. Read the status board the agent prints after each step, which shows each subtask's state and retries.

• Name the deliverable and the number of items up front.
• Ask for verifiable criteria, such as a required output format or a set of fields per item.
• Watch the status board for FIX or ESCALATED entries.