zcyynl/claw-multi-agent

Overview

This skill runs multiple AI agents in parallel to treat a single task like a coordinated team. It orchestrates researchers, analysts, writers and coders to perform simultaneous research, multi-model comparisons, and pipeline workflows—often cutting elapsed time by roughly 50–65%. The skill enforces a predictable interaction pattern: announce activation, present a short plan, spawn agents, consolidate results, and deliver a polished report.

How this skill works

The skill analyzes the incoming request and dynamically chooses agent count, roles, and models, then spawns agents either in parallel (same round) or sequentially when dependencies exist. Parallel agents run independently with tools (search, file I/O, code execution) and return short summaries; the main agent consolidates and rewrites into structured output. It supports three modes—Orchestrator (tool-enabled parallel), Pipeline (pure-text parallel/sequential), and Hybrid (search then multi-draft)—and always saves a report file before delivery.

When to use it

• Research multiple topics or competitors simultaneously to save time
• Compare answers from different LLMs or prompt styles in parallel
• Batch-process many documents: translate, summarize, or analyze at scale
• Run a code pipeline (plan → code → review) with automated hand-offs
• Generate several draft versions or angles for editorial selection

Best practices

• Always send an immediate activation announcement before any work begins
• List each agent in the pre-spawn plan with emoji, role, explicit model, and one-line task
• Spawn dependent agents only after their inputs are available (parallel then sequential rounds)
• Limit sub-agent outputs to concise summaries; let the main agent synthesize into a final report
• Save a .md report file first and choose delivery by channel (attachment, doc link, or message)

Example use cases

• Run 5 researchers to research five frameworks in parallel and a sixth analyst to compare results
• Spawn Claude, Gemini and Kimi simultaneously to compare model outputs on the same prompt
• Process 50 documents in parallel to extract key findings and save a combined report
• Execute a pipeline: planner spawns a coder, then a reviewer runs after the coder returns
• Generate three stylistic drafts of a product description in parallel for A/B selection