token-economicslisted

# Token economics On a metered plan this is money. On a managed Enterprise seat it is **quota and admin-set spend caps** — the constraint arrives as a limit you hit mid-close rather than an invoice, which is worse, because it hits when you are busiest. The rule that governs everything below: > **Never put data into context that a script could reduce first.** Finance work is unusually exposed here because the artifacts are large — GL extracts, aging reports, transaction detail. A single careless read of a full-year GL can consume more context than an entire close's worth of reasoning. **Measure before optimizing.** Run `usage-audit` first; this skill is the doctrine, that one is the instrument. Optimizing from intuition is how people spend an afternoon saving nothing. --- ## The data-tier hierarchy Applies to every "look at this data" task. **Try tiers in order and stop at the first that works.** | Tier | Method | Typical context cost | |---|---|---| | **1** | **Aggregate at source.** `GROUP BY` in the warehouse; return the 40 rows you need | tens of tokens | | **2** | **Script it locally.** Python reduces the file and prints distilled JSON | ~100-300 tokens | | **3** | **Targeted read.** `Grep`/`head` for the specific rows, never the whole file | hundreds | | **4** | **Subagent.** Delegate the bulk read; only the conclusion returns to your context | the summary only | | **5** | **Read it into context.** Only when the task genuinely needs every row | thousands to tens