ultrawork

merozemory

•

1933 months ago

Source Raw

Maximum parallel execution mode with droid orchestration for high-throughput task completion

Quick Install

npxskills add merozemory/oh-my-droid--skill ultrawork

Instructions

Ultrawork

Orchestrate parallel sub-agents across model tiers to complete multi-part tasks faster and cheaper.

Workflow

Follow these steps in order for every ultrawork session:

Decompose -- Break the user request into independent sub-tasks. List each task with its domain (Analysis, Execution, Search, etc.).
Route -- Assign each sub-task to the cheapest sufficient agent tier using the routing table below (LOW first, escalate only when needed).
Dispatch -- Launch independent sub-tasks in parallel via Task(). Set run_in_background: true for any command expected to take longer than a few seconds (builds, installs, test suites, Docker operations).
Monitor -- Track each sub-agent to completion. If a sub-task fails or returns incomplete results, escalate to the next tier and retry.
Verify -- Run the verification checklist (below). If any item fails, continue working until all pass.
Clean up -- Delete .omd/state/ultrawork-state.json so no stale state persists.

Smart Model Routing

Always start at the lowest tier that can handle the task. Escalate only on failure or when complexity demands it: LOW (Haiku) -> MEDIUM (Sonnet) -> HIGH (Opus).

Available Agents by Tier

Domain	LOW (Haiku)	MEDIUM (Sonnet)	HIGH (Opus)
Analysis	`architect-low`	`architect-medium`	`architect`
Execution	`executor-low`	`executor`	`executor-high`
Search	`explore`	`explore-medium`	-
Research

Tier Selection Guide

Task Complexity	Tier	Examples
Simple lookups	LOW	"What does this function return?", "Find where X is defined"
Standard work	MEDIUM	"Add error handling", "Implement this feature"
Complex analysis	HIGH	"Debug this race condition", "Refactor auth module across 5 files"

Routing Examples

OMD auto-injects the correct model for built-in tiered agents when missing.

If you do specify model, use full model IDs (e.g. claude-haiku-4-5-20251001), not shorthands.

// Simple question → LOW tier (saves tokens!)
Task(subagent_type="oh-my-droid:architect-low", model="claude-haiku-4-5-20251001", prompt="What does this function return?")

// Standard implementation → MEDIUM tier
Task(subagent_type="oh-my-droid:executor", model="claude-sonnet-4-5-20250929", prompt="Add error handling to login")

// Complex refactoring → HIGH tier
Task(subagent_type="oh-my-droid:executor-high", model="claude-opus-4-5-20251101", prompt="Refactor auth module using JWT across 5 files")

// Quick file lookup → LOW tier
Task(subagent_type="oh-my-droid:explore", model="claude-haiku-4-5-20251001", prompt="Find where UserService is defined")

// Thorough search → MEDIUM tier
Task(subagent_type="oh-my-droid:explore-medium", model="claude-sonnet-4-5-20250929", prompt="Find all authentication patterns in the codebase")

Background vs Foreground Execution

Set run_in_background: true for any command that typically takes more than a few seconds:

Background (`run_in_background: true`)	Foreground (blocking)
`npm install`, `pip install`, `cargo build`	`git status`, `ls`, `pwd`
`npm run build`, `make`, `tsc`

Verification Checklist

Before stopping, every item must pass. If any fails, continue working.

Tasks complete -- Zero pending or in-progress sub-tasks remain
Functionality -- All requested features work as specified
Tests -- All tests pass (when applicable)
Errors -- Zero unaddressed errors in output

State Cleanup

On completion, delete state files -- do not leave them with active: false:

rm -f .omd/state/ultrawork-state.json

Loading…

Related Skills

skill-creator

Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.

367,360