AI 코딩 비용 줄이는 지점 찾는 프롬프트

Find where the cost is going and what to change.

Check in this order — this is ordered by how much it usually saves:
1. **Repeated failed attempts.** The same task run five times because the first four missed. This is usually the largest single line and is a prompt problem, not a pricing problem.
2. **Context being resent.** Whole files or histories included when a fragment would do. What is in context that does not need to be.
3. **Rework** — work produced, rejected, redone. Where an upfront question would have prevented it.
4. **Task fit.** Work being given to a large model that a smaller one, a script, or a search would do. Mechanical edits are the common case.
5. **Only then, model choice.** *Downgrading before fixing the above usually increases total cost, because worse output means more attempts.*

Per finding: what to change, and roughly what share of current spend it touches.

Then:
- The single change with the largest saving.
- What to measure to know whether it worked.
- Where spending more is correct — a task worth doing well once beats three cheap attempts.

Rules:
- *Do not state current prices or rate limits as fact.* Work from the usage I gave you; if I gave none, say what to collect.
- Do not recommend cutting usage that is producing value. The goal is removing waste, not doing less.