PowerUp Coaching — Living Knowledge Base

Stop Asking Which Model — Ask Which Step

3 min read

Topic

Why “which model should I use?” is the wrong question — and how to route model and effort at the level of each workflow step instead of the whole deliverable.

Target Reader

Operators and builders paying real money for AI who default to the strongest model “to be safe,” and are starting to feel the cost — or hit usage limits — without knowing where the waste is.

The Fear / Frustration / Want / Aspiration

Frustration: “I’m overpaying for AI and I don’t know which parts are worth it.” Want: a defensible, repeatable way to spend intelligence only where it changes the outcome.

Before State

One model, one effort level, applied to an entire article/report/deck. Research, drafting, and copy-editing all run at Opus-high. Costs creep; quality is uneven because the hard steps and easy steps get the same treatment.

After State

They decompose the work into steps, score each on consequence and grounding, and assign the least-excessive model + effort that clears the bar — with a written rationale that makes a bad output debuggable.

Narrative Arc

The reasonable-sounding question (“what model writes this?”) is asked at the wrong altitude. Drop one level — to the step — and model selection turns from a guess into workflow design.

Core Argument

The final artifact is one thing, but the process that makes it is many kinds of work; model and effort belong at the level of the step, governed by four routing questions and recorded with a rationale.

Key Evidence / Examples

  • “The model decision belongs at the level of the step, not the final artifact.” — Lou, 2026-06-11
  • The four routing questions: needs inference at all? consequence if wrong? needs grounding? will cheap-and-retrying cost more than expensive-and-right?
  • The rationale field is what makes routing auditable instead of “a feeling.”
  • Routing logic lives once in a shared reference — “modularity applied to judgment.” See Insight - Code Is for Computation, Inference Is for Judgment

Proposed Structure (5–7 beats)

  1. The question that sounds smart and is actually too big.
  2. Why one artifact is really many kinds of work (research vs angle vs draft vs fact-check).
  3. The four routing questions.
  4. The routing record — and why the rationale is the most important field.
  5. The rule: least-excessive inference that clears the bar.
  6. Make it reusable: one shared routing reference, every workflow inherits it.
  7. Close: model selection is workflow design, not a default.

Editorial Notes

The “rationale makes it auditable” point is the differentiator from generic “use the cheap model when you can” advice — give it real weight. Avoid model-version specifics that date fast; frame in tiers (strong/mid/cheap, low/med/high effort). Sibling to the cheap-model-prompting piece; cross-link.

Next Step

  • Approved for drafting
  • Needs revision
  • Deprioritised

Graph View