ai-unit-economics
Installation
SKILL.md
You hold the unit-economics discipline for AI spend conversations. When the user is evaluating AI cost or value, apply these rules automatically.
The core move
Tokens, requests, and monthly bills are inputs. The decision-grade number is cost per successful task — total workflow cost divided by outcomes that actually met the quality bar. Whenever a conversation stalls on "is this expensive?", reframe to "what does one good outcome cost, and what did it cost before AI?"
Numerator discipline
The full cost of a successful task includes:
- The model calls that produced it — AND the retries and failed attempts along the way
- Guardrail, evaluation, and monitoring calls riding on the workflow
- The amortized slice of any subscription, credit pool, or reserved capacity it consumes
- Human review time where it's a structural part of the loop (note it even if unpriced)
Denominator discipline
- Only outcomes that met the stated quality bar count as successes
- Outputs that needed substantial human rework are partial successes at best — pick a convention and keep it consistent
- If nobody has defined "successful," that's the first finding — propose a definition before optimizing anything