ultrawork_skill

Overview

This skill is a parallel execution engine that launches multiple agents concurrently to complete independent tasks faster while controlling model routing and token costs. It is a composable component that provides parallelism and smart tier selection but does not handle persistence, deep verification loops, or long-term state. Use it when you want high-throughput task completion by delegating work across appropriate model tiers.

How this skill works

Ultrawork classifies incoming work into independent and dependent tasks, routes each job to an appropriate agent/model tier, and fires all independent tasks simultaneously. Long-running operations are dispatched to background execution while quick checks run in the foreground. It performs lightweight verification when invoked directly and delegates persistence or full verification to higher-level modes that include ultrawork.

When to use it

• You have multiple independent tasks that can run at the same time
• User explicitly requests ultrawork, ulw, or parallel execution
• You need to delegate work to multiple agents concurrently to save wall-clock time
• Tasks benefit from concurrent execution but the caller will manage completion and state
• You want explicit control over model tier routing to optimize cost and quality

Best practices

• Classify tasks into parallel-safe and dependent groups before launching work
• Always pass an explicit model parameter when delegating to an agent
• Consult the agent-tier reference before selecting models (LOW/MEDIUM/HIGH)
• Run long installs, builds, and test suites with run_in_background: true
• Execute quick file reads or status checks synchronously in the foreground
• Verify builds/typechecks and affected tests after parallel work completes (lightweight checks)

Example use cases

• Run three independent code fixes/tests simultaneously on different model tiers to reduce turnaround time
• Launch a package install and build in background while running small documentation edits in foreground
• Dispatch standard implementation tasks to MEDIUM tier and complex refactors to HIGH tier in parallel
• Fire multiple independent integration tests at once, then run a short verification pass
• Delegate simple type or formatting fixes to LOW tier agents concurrently with larger feature work