askill
ultrawork

ultraworkSafety 100Repository

This skill should be used when the user has a large, complex task requiring: (1) multi-file or multi-step implementation, (2) major refactoring or architecture changes, (3) new feature spanning multiple components. This skill should NOT be used for simple fixes, single-file edits, or quick questions. Activates strict verification with planning, success criteria, and evidence collection.

2 stars
1.2k downloads
Updated 2/22/2026

Package Files

Loading files...
SKILL.md

Ultrawork Mode

Overview

Ultrawork enforces verification-first development - a strict mode where every completion claim requires concrete evidence.

Core Principles

  1. No implementation without planning - Explore codebase, design approach, decompose into tasks
  2. No completion without evidence - Every criterion needs proof (command output, test results)
  3. No partial work accepted - "Should work" or "basic implementation" triggers automatic rejection

Workflow Phases

PLANNING → EXECUTION → VERIFICATION → COMPLETE
    ↑                        │
    └── (Ralph Loop on fail) ←┘
PhaseDescriptionKey Outputs
PlanningExplore → Design → Task decompositiondesign.md, tasks/*.json
ExecutionWorkers implement tasks in parallel wavesCode changes, evidence
VerificationVerifier audits ALL evidence against criteriaPASS/FAIL decision

Activation

Start ultrawork with the /ultrawork command:

/ultrawork "your goal"         # Interactive: asks questions, user approves plan
/ultrawork --auto "your goal"  # Auto: decides autonomously, no confirmations
/ultrawork --plan-only "goal"  # Planning only, no execution

Mode Selection

ModeBest ForUser Interaction
Interactive (default)Complex features, unclear requirementsQuestions + approval
Auto (--auto)Well-defined tasks, CI/CDNone
Plan-only (--plan-only)Design review before implementationPlanning only

When to Use

USE ultrawork for:

  • Multi-file implementations (3+ files)
  • Architecture changes or refactoring
  • New features with design decisions
  • Work requiring verification trails

DON'T use ultrawork for:

  • Single-file edits or bug fixes
  • Documentation updates
  • Quick questions or exploration
  • Tasks completable in < 5 minutes

Zero Tolerance Rules

Verification automatically FAILS if output contains blocked patterns.

See references/blocked-patterns.md for complete list.

Common blocked phrases:

  • "should work", "probably works"
  • "basic implementation", "simplified version"
  • "TODO", "FIXME", "not implemented"

Evidence Requirements

Every completion claim requires concrete proof.

ClaimRequired Evidence
"Tests pass"npm test output with exit code 0
"Build succeeds"Build command output with exit code 0
"Feature works"Demo, test output, or screenshot
"Bug fixed"Before/after comparison

See references/verification-protocol.md for detailed verification requirements.


Session State

Each session has isolated state in ~/.claude/ultrawork/sessions/${CLAUDE_SESSION_ID}/.

See references/state-schema.md for complete schema documentation.

Key files:

  • session.json - Phase, options, metadata
  • context.json - Exploration summaries
  • exploration/*.md - Detailed findings
  • tasks/*.json - Task definitions and evidence

Related Commands

CommandPurpose
/ultrawork-statusCheck current phase and progress
/ultrawork-evidenceView collected evidence log
/ultrawork-cleanClean up sessions (interactive/all/stale)
/ultrawork-execExecute existing plan
/ultrawork-planPlanning phase only

Additional Resources

Reference Files

  • references/blocked-patterns.md - Complete list of blocked phrases and why
  • references/verification-protocol.md - Detailed verification requirements
  • references/state-schema.md - Session and task JSON schemas

Install

Download ZIP
Requires askill CLI v1.0+

AI Quality Score

58/100Analyzed 2/24/2026

A well-structured methodology guide for a verification-first development workflow. Provides clear activation commands, workflow phases, and usage criteria. However, it lacks actionable step-by-step implementation and heavily relies on external reference files that may not exist. The skill is tailored to a specific Claude agent environment (internal_only), which limits reusability. The tags are mismatched with actual content. Score benefits from clear structure and 'when to use' section, but penalized for external file dependencies and lack of practical implementation guidance.

100
75
45
60
55

Metadata

Licenseunknown
Version-
Updated2/22/2026
Publishermnthe

Tags

ci-cdgithub-actionsllmtesting