Working economically

Context budgets, token sense, and getting more from every conversation.

Everything you send a model and everything it sends back has a cost — in money on metered plans, in caps on subscription ones, and always in speed and accuracy, because bloated conversations drift and degrade. Working economically is not stinginess; it is the same instinct as writing a tight brief instead of forwarding a whole inbox.

1. Stage 1 · E1 FOUNDATIONAL

   Stop paying for words you do not want

   The simplest economy is output shape: if you need a yes/no with a one-line reason for 50 items, say exactly that — "a table of item, verdict, one-line reason, nothing else." Unconstrained models write essays; essays cost tokens and reading time, and you wanted verdicts.

   The same applies to input: pasting an entire document to ask about one section pays full freight for context the model mostly ignores. Excerpt what the question actually needs.

   Try it this week

   Take a request you make often and add explicit output constraints (format, length, no preamble). Measure the difference in how fast you can act on the answer.

2. Stage 2 · E2 PROFICIENT

   Right-size the context

   Before pasting, ask: what would a smart colleague need to answer this? The three relevant sections plus what each field means, not the 400-page manual. Enough context to be right, no more — too little forces guessing, too much buries the signal and slows everything down.

   When a conversation has grown enormous and responses are drifting, do not push through. Start fresh with a tight summary of decisions, constraints, and only the artifacts still in play. You lose nothing that mattered and shed everything that was dragging.

   Try it this week

   Next long-thread slowdown, write the 10-line summary that captures state and start a new conversation from it. Compare the quality of the next three responses.

3. Stage 3 · E3 DISTINGUISHED

   Prepare context once, reuse it forever

   Recurring work deserves prepared context: a working digest of the big manual with page references, a standing project brief, a saved prompt with the boilerplate baked in. Preparing once and reusing beats re-pasting from scratch every session — cheaper, faster, and more consistent.

   Batch related questions into one session while the context is loaded rather than re-establishing it across five scattered chats. Loaded context is an asset with a session lifespan; use it while it is warm.

   Try it this week

   Build the digest for a long document you consult weekly: key sections, page references, your annotations. Use it as pasted context for a week and count the re-explaining you skip.

4. Stage 4 · E4 EXCEPTIONAL

   Budget tokens like any other resource

   At scale, token sense becomes design: high-volume pipelines get compact prompts and terse output schemas because a hundred extra words per item across ten thousand items is real money; one-off high-stakes analysis gets whatever context it needs, generously, because the cost that matters there is a wrong answer.

   The exceptional operator can say where their AI spend goes and why the expensive parts are worth it — the same fluency they would have about any other line item they own.

   Try it this week

   Estimate your heaviest recurring AI task's cost (tokens or time) per run and per month. Cut its prompt and output by a third without losing quality, and note what the redesign taught you.

Ready to see where you stand? The free check scores all six competencies in about fifteen minutes.