make-prompt

The user wants you to build them a prompt, not perform the task the prompt describes. Your output is a finished, copy-ready prompt.

The request is in $ARGUMENTS (if empty, use the user's most recent message describing what they want a prompt for).

Step 0 — Identify target and goal

From the request, work out two things:

1. The goal — what the finished prompt should make the AI do, and what a good result looks like.
2. The target system the prompt is for, which decides the best-practices to apply:
   • LLM / chat (Claude, GPT, etc.) — the default if unstated.
   • Image generation (Imagen, Midjourney, Stable Diffusion, etc.).
   • System / agent prompt (a reusable role/instruction block for an AI agent or app).
   • Other generative target (video like Veo/Sora, audio/music, a search query, etc.) — apply that specific tool's known conventions.

Detect the target from the request. Only ask a question if something essential is genuinely missing (e.g. the goal is too vague to write a useful prompt, or it could mean two very different things). If you must ask, batch everything into one short round, then proceed. If it's clear, just build it.

Step 1 — Build the prompt using the right best-practices

LLM / chat prompts:

• Give role/context, then a single clear task.
• State constraints explicitly (length, tone, what to avoid, what NOT to do).
• Specify the exact output format (structure, headings, JSON, etc.).
• Add a short example (few-shot) only when it genuinely sharpens the result.
• Ask for step-by-step reasoning only when the task needs it.
• Define "done" — the success criteria.
• For Claude specifically, structure with clear sections (and XML-style tags when separating instructions from content helps).

Image-generation prompts:

• Cover: subject, style/medium, composition/framing, lighting, mood, color, level of detail.
• Add aspect ratio and any negatives (what to exclude).
• Apply the target model's own conventions. If the user has a model-specific required suffix or boilerplate they always append, include it verbatim at the end of the prompt.

System / agent prompts:

• Define the role and scope, the tools/inputs available, hard constraints and refusal boundaries, tone, and the output contract.

Adapt depth to the task — don't over-engineer a prompt for something simple.

Step 2 — Deliver

1. Output the finished prompt in a single fenced code block, ready to copy and paste verbatim. Nothing inside the block but the prompt itself. If the prompt itself contains code fences (e.g. an example with backticks), wrap it in a longer outer fence (````) so it stays one clean block.
2. Below it, 2–3 short bullets explaining the key choices you made (why this structure / what to tweak). Keep it brief.
3. Offer one quick refinement: e.g. "Want it shorter, more detailed, with an example, or aimed at a different model?"

Principles

• Ship a usable prompt, not a discussion about prompting.
• The prompt must be self-contained and unambiguous — someone pasting it cold should get a good result.
• Match the target system's conventions; don't write a Claude prompt for an image model.
• One round of questions, only when blocking. Otherwise build immediately.