Paste a prompt or the outline of an LLM workflow, plus your call volume and the price per million tokens you pay, and get a plan to cut the bill without hurting quality. It shows where the tokens go, then lists the savings in order of size, each with the quality risk and how to check it: shorter instructions, putting the unchanging part of the prompt first so caching works, sending small tasks to a cheaper model, capping output length, trimming retrieved context and batching. It gives a rewritten prompt, the arithmetic for monthly savings using your own prices, and a way to measure the result.