A former Fortune 500 Chief Data Officer has released an open-source Claude skill designed to flip how development teams think about measuring AI-assisted delivery value. Harveer Singh, who previously led data operations at Truist Bank and Western Union before founding Rizz Wireless, published the "Earned vs Burned" framework on GitHub this week with a blunt message: most teams are tracking metrics that don't matter.

The Problem With Current Delivery Metrics

Singh argues that standard development measurements—token usage, lines of code written, story points closed, hours logged—are fundamentally "burned metrics." They measure effort expended rather than value delivered. A model that hallucinates in 100 tokens isn't producing better outcomes than one solving the same problem correctly in 10,000 tokens. Code that never reaches production earns nothing regardless of how elegant it is. The framework's core thesis: until a tangible, verifiable business outcome is achieved, every hour and every token spent on a task is simply burned.

A Five-Level Hierarchy for Outcome Tracking

The Earned vs Burned system maps delivery progress across five levels with corresponding "earn" values. Level 0 (Not Started) earns nothing. Level 1 (In Progress) still shows zero earned despite accumulating effort. Code written and unit-tested reaches Level 2 (Dev Complete) at 25% earn value. QA acceptance pushes work to Level 3 at 60%. Deployed to production sits at 85%, with full credit (100%) reserved exclusively for Level 5: Outcome Verified, where KPIs have moved, users confirm value, or revenue has been impacted.

The Metric AI Shops Actually Need

Perhaps the most interesting innovation is "Earned per AI Token"—a ratio Singh claims the industry currently lacks entirely. Rather than tracking raw token consumption, teams would monitor whether each unit of AI spend generates increasing value over time. Other targets include an Earn Rate percentage above 70% (L5 outcomes divided by total stories), an Earned/Hours ratio exceeding 0.10, and climbing Outcome Verification percentages as a culture indicator.

Integration and Installation

The skill pulls tasks from Linear, Asana, GitHub Issues, Jira, or Azure DevOps, scoring each against the hierarchy and generating what Singh calls an "Earned Value Report—one page, three numbers, replaces your velocity report." Teams can also paste task lists or upload CSVs. Installation works via Claude Desktop's skill management or manual configuration for Claude Code users by adding the path to settings.json.

Beyond Internal Teams

The framework explicitly targets outsourced and offshored delivery scenarios, where Singh sees traditional metrics as particularly inadequate for proving value to clients. Vendor teams can use outcome-based scoring to demonstrate concrete business impact rather than hours logged or tokens consumed—addressing a real pain point in managed services relationships.

Key Takeaways

  • Most development metrics measure effort (burn) not outcomes (earned)
  • The framework reserves full credit only when verified business outcomes are achieved
  • "Earned per AI Token" replaces raw token tracking with value efficiency measurement
  • Works for internal teams, vendors, or hybrid arrangements across multiple project tools
  • Skill code is MIT licensed; the framework methodology remains Singh's intellectual property

The Bottom Line

This is the kind of practical tooling that gets ignored until you need it and can't find anything like it. If you're running AI-assisted delivery at any scale, measuring anything other than verified outcomes is just lying to yourself with extra steps.