Blog · Product / Eng · 20 Aug 2026

How Do Compact JSON Schemas Cut Prompt Cost?

Every extra key name you send back into the next agent turn is billed input. Compact keys, a 1-token model id, and a named umbrella cut that overhead and reduce retries. ConvoMargin then measures whether margin actually moved.

Why do short JSON keys matter for LLM spend?

Agents often re-inject the last result as context. A verbose log ("Session using Cursor Grok 4.6 with web search…") is readable. It is also token-heavy. A receipt like {"m":"grok-4.6","ti":800,"to":400,"c":0.01,"rev":0.5,"mg":0.49} is the same P&L in fewer tokens. Daniel's Corner already uses this for private taxonomies: compact keys, lowercase, alias map, then a 1-token canon.

ConvoMargin applies the same rule to cost tracking. Model nicknames collapse: "Cursor Grok 4.6" and "grok" both become grok-4.6. Tools collapse: internet, search, browse become web. You can roll up per-prompt mix without a new product category.

TLDR: Compact keys plus alias maps cut re-injected tokens. Try a receipt.

What is an umbrella before the first prompt?

LLMs are probability engines over word chunks. If you do not name the topic cluster, the first prompt wanders and you pay for wrong branches. An umbrella is a one-line topic plus 8–15 core chunks the answer must use. Edge chunks only when the last run missed intent. That is the Daniel's Corner prompt-umbrella file, pointed at ConvoMargin: SaaS P&L, enterprise ROI, support tickets, model mix, schema.

Umbrella prompting is not margin math. It is a way to spend fewer tokens getting to a usable answer. The receipt still logs whichever model ran. You can switch Grok, Sonnet, or a cheaper fallback on the next turn. The next row just has a different m.

TLDR: Name the cluster first. Log the model that ran. See umbrellas.json.

How does ConvoMargin measure the saving?

The preview estimates compact vs verbose tokens with chars/4 and prices the difference at your selected tier input rate, times interactions per day. That number is schema overhead, labeled modeled. It is not a Cursor invoice. An Audit Sprint can compare your real logs against the compact receipt shape. Waitlisted middleware is meant to emit this row at the API boundary.

Public schema: prompt-receipt.json. Worked example: receipts-aug20.json (this site's Grok 4.6 sessions).

TLDR: Measure schema overhead, then measure conversation margin. Book an audit.