orchestrate_skill

Overview

This skill helps you orchestrate multi-agent work at scale by recommending orchestration strategies, prompt structures, and background vs foreground rules. It codifies patterns for fan-out research swarms, parallel feature builds, wave dispatch, sequential pipelines, and multi-dimensional audits. Use it to reduce conflicts, speed delivery, and adapt review overhead as trust grows.

How this skill works

The skill inspects the type of work, codebase dependencies, and desired throughput to recommend one of six strategies (Research Swarm, Epic Parallel Build, Sequential Pipeline, Parallel Sweep, Multi-Dimensional Audit, Full Lifecycle). It outputs concrete prompt templates, dispatch cadence rules, partitioning guidance, and git hygiene instructions so agents can run in parallel without stepping on each other. It also provides a foreground/background decision matrix and a trust-gradient for review overhead.

When to use it

• Kicking off large research or SOTA investigations (Research Swarm)
• Implementing many independent features across modules (Epic Parallel Build)
• Working on dependent tasks or shared files that need review gates (Sequential Pipeline)
• Applying the same transformation across multiple directories or modules (Parallel Sweep)
• Running comprehensive reviews from multiple perspectives before release (Multi-Dimensional Audit)
• Starting a greenfield project and needing an end-to-end session plan (Full Lifecycle)

Best practices

• Partition work by directory/module and give each agent an explicit output path to avoid overlap
• Prefer background agents for independent research or isolated builds; use foreground for dependent tasks and reviews
• Provide explicit git instructions: stage only created files, avoid git add ., and let orchestrator handle pushes
• Dispatch in waves and monitor results; use short delays between agents and longer gaps between waves for gap analysis
• Adopt a trust gradient: heavy review early, reduce ceremony as patterns prove reliable

Example use cases

• Run a 30-agent research swarm to produce a knowledge corpus for technology selection
• Dispatch 20+ feature agents in waves, each implementing a single epic in its own module directory
• Execute a 5-step sequential pipeline for integration boundary changes with implement->review->fix gates
• Fan-out 6 agents to add type annotations across packages, one agent per package (Parallel Sweep)
• Launch 6 specialized reviewers (security, perf, tests, spec, integration, quality) simultaneously and synthesize findings into prioritized fixes