prompt-me

Prompt-Me: Refine a Prompt With a Small Model First

Turn a rough request into a precise, self-contained prompt for a powerful coding model — but do the refinement work on a small, cheap model and surface at most 3 clarifying questions to the human. The expensive model then gets a complete spec on its first attempt instead of discovering gaps mid-task.

Core principle: Spend cheap tokens to save expensive ones. Analysis, restructuring, and question-finding are mechanical scaffolding — offload them to a small model so the large model only does the hard engineering.

When to Use

• A request is vague, one-line, or under-specified ("add auth", "make it faster")
• You're about to hand a task to an expensive/capable model and want it right first try
• You want to cut iteration cycles with a large model
• The human explicitly asks to improve, sharpen, or "prompt me" on a task

When NOT to use: the request is already a complete, unambiguous spec, or the task is trivial and any reasonable interpretation is fine.

The Loop

digraph prompt_me {
    "Capture raw request + context" [shape=box];
    "Dispatch SMALL model with prompt-engineer system prompt" [shape=box];
    "Model returns: refined prompt + assumptions + <=3 questions" [shape=box];
    "Any questions?" [shape=diamond];
    "Ask human (ask_user, one at a time)" [shape=box];
    "Fold answers into final prompt" [shape=box];
    "Deliver final prompt (offer to run it)" [shape=doublecircle];

    "Capture raw request + context" -> "Dispatch SMALL model with prompt-engineer system prompt";
    "Dispatch SMALL model with prompt-engineer system prompt" -> "Model returns: refined prompt + assumptions + <=3 questions";
    "Model returns: refined prompt + assumptions + <=3 questions" -> "Any questions?";
    "Any questions?" -> "Ask human (ask_user, one at a time)" [label="yes"];
    "Any questions?" -> "Deliver final prompt (offer to run it)" [label="no"];
    "Ask human (ask_user, one at a time)" -> "Fold answers into final prompt";
    "Fold answers into final prompt" -> "Deliver final prompt (offer to run it)";
}

Steps

1. Capture the raw request and any cheap context. Take the human's words verbatim. Add only context you already have (repo stack, key files, constraints). Don't go investigate the whole codebase — that's the large model's job.

2. Dispatch the small model. Use the task tool with agent_type: "general-purpose" and model: "gpt-5.4-mini" (or another small, cheap model). Pass the contents of prompt-engineer-system-prompt.md as the dispatch prompt, substituting {{RAW_REQUEST}} and {{CONTEXT}}. The small model returns three sections: REFINED PROMPT, ASSUMPTIONS, QUESTIONS.

3. Ask the human the questions. Take the (0–3) questions and ask them with the ask_user tool — one question per call, using the model's multiple-choice options as choices (put the (recommended) one first). If the model returned no questions, skip straight to delivery. Never exceed 3 questions.

4. Fold answers back in. Update the refined prompt with the answers and resolve the matching assumptions. For small edits do it yourself; if answers change the shape of the task significantly, re-dispatch the small model with the answers added to {{CONTEXT}}.

5. Deliver. Present the final prompt in a copy-pasteable block. Offer to run it directly with the powerful model, or hand it to the human to use.

Quick Reference

Thing	Default
Helper model	gpt-5.4-mini (any small, cheap model)
Helper agent_type	general-purpose
Max questions to human	3 (often 0–1)
Question delivery	ask_user, one at a time, with options
Final prompt sections	Objective, Context, Requirements, Constraints, Acceptance Criteria, Deliverables, Out of Scope

Common Mistakes

• Refining on the expensive model yourself. Defeats the purpose — dispatch the small model to do the analysis.
• Interrogating the human. More than 3 questions, or questions whose answers you could reasonably assume. Cap at 3; prefer 0–1.
• Skipping the questions entirely. If the small model surfaced a real unknown, ask it — that's the iteration you're trying to save the large model from.
• Letting the small model implement the task. It only refines the prompt.
• Inventing codebase facts. Assumptions must be labeled and surfaced, not presented as certainty.

Files

• prompt-engineer-system-prompt.md — the system prompt to hand the small model.