meta prompt
Cheap-Model Prompt
Cost Optimization / Systems — use the smartest model to prompt the cheapest model
Use it for
Inverts the usual model-selection instinct. Instead of running a task on a premium model, you fix the cheapest viable model and hire the premium model for one job: author a prompt that compensates for the cheap model’s limits so it performs like the expensive one. The reasoning is paid for once (in the prompt) and reused on every cheap inference after. Use it for any repetitive task where the reasoning is stable — especially forked/spawned sub-steps where you already pin the child’s model.
The prompt
> I'm going to run the task below on **[CHEAP MODEL — e.g. Haiku 4.5]** at **[effort: low/medium/high]**.
>
> You are **[STRONG MODEL — e.g. Opus 4.8]**. Knowing the specific capabilities and constraints of [CHEAP MODEL], write a single prompt that gets [CHEAP MODEL] to perform this task **as well as you would**. Bake in:
> - the strategy or method it should follow,
> - exactly where it should think step-by-step,
> - any sub-step it's likely to get wrong without explicit instruction,
> - the output format and quality bar.
>
> Don't solve the task yourself — produce the prompt I'll hand to [CHEAP MODEL].
>
> **Task:** [DESCRIBE THE TASK + paste any inputs/examples]
---Worked example
No worked example recorded yet.