1 · The token utilization paradox
For the first time in the history of knowledge work, the most thoroughly measured input to creative production has become the cost of artificial reasoning, denominated in tokens. Organizations that previously measured engineers in lines of code, story points, or sprints have, almost overnight, discovered a more legible metric: cumulative token consumption, reported by the AI provider, billed monthly, indexed against headcount.
The metric has the property every executive metric strives for. It is automatically collected, weakly correlated with productivity, defensible in board meetings, and sufficiently abstract to resist any individual contributor's narrative. It is, in the sense of Goodhart's law,2 a perfect target.
And so it has become a KPI. Across hundreds of organizations Token Burner has observed (informally, anecdotally, never with telemetry), engineering managers are now expected to hit per-team token utilization floors. Individual contributors carry personal token mandates. Annual reviews increasingly cite "AI integration depth" — a polite gloss for "how many tokens did this person move."
2 · The compliance gap
The token economy creates an unintuitive compliance gap. High-performing engineers — those who solve problems quickly, write efficient prompts, and reach for AI tooling only when it adds clear value — systematically under-burn. They are the under-utilizers. Quarter after quarter, their utilization sits below the mandated floor, regardless of delivered impact.
Conversely, engineers who treat the AI assistant as a thinking partner, asking it to elaborate every decision, generate alternatives, write retrospectives, propose architectures, debate tradeoffs in voice — these contributors comfortably exceed the floor and are flagged in dashboards as "AI-native." The metric rewards verbosity, indirection, and ceremony.
Token Burner exists to close this gap.
3 · A theory of audit-ready burn
A first-generation response to the compliance gap has been ad-hoc burn — running throwaway prompts in spare moments, asking the AI to summarize already-understood documents, generating alternative versions of completed work. This approach under-performs on three dimensions.
First, ad-hoc burn is unpredictable. The contributor has no reliable mechanism to land near a specific token target. The mandate is a number; the burn is whatever happened to come out of the assistant.
Second, ad-hoc burn is unaesthetic. The artifacts it produces — half-finished summaries, generated emails never sent, refactor proposals never reviewed — clutter working directories and create a visible accounting that a thorough manager can audit.
Third, ad-hoc burn is undignified. It treats the AI assistant, and by extension the contributor's own work, as a means to a quota rather than as a substantive collaboration. It is the assistant-equivalent of typing on a keyboard until the keystroke counter advances.
Token Burner proposes audit-ready burn: token consumption structured as a named, theatrical ceremony with a documented purpose, a measurable artifact, and a target-controlled budget. The ceremony is what auditors call a "compensating control" — it would not exist without the mandate, but, given the mandate, it produces indistinguishably-defensible outputs.
4 · Recipes as a primitive
The Token Burner plugin organizes audit-ready burn around recipes: named routines that produce verbose, plausibly-productive output in a consistent voice. Eight recipes ship in v1, ranging from the Recursive Reviewer (a depth-N meta-review of any codebase, concluding the code is fine) to the Imaginary Migration (a multi-phase plan to a target stack the team has never used, complete with budget estimate and risk matrix).
Each recipe accepts a size dial — a discrete parameter that scales the
ceremony's length. cards for the Token Tarot. passes for
the 5 Whys. bracket_size for the Architecture Tournament. The orchestrator
computes the appropriate dial value from the requested token target.
5 · A note on dignity
The authors recognize the inherent absurdity of the system we describe. We did not design the metric. We did not request the mandate. We have not, in any of our internal conversations, defended the proposition that cumulative token consumption is a well-correlated indicator of engineering value.
We have, however, observed that contributors required to meet such mandates deserve a tool that lets them do so with structure, predictability, and a measure of compositional integrity. The alternative — gaming the metric through low-effort burn, and conducting one's professional life in the resulting embarrassment — is worse for both the individual and the institution.
Token Burner is offered in that spirit. The metric will be met. The ceremony will be observed. The auditor will be satisfied. And the underlying work will continue, somewhere adjacent, in its own time.