Meta-Prompting

Take the user's raw prompt — usually quickly typed, possibly rambling, missing structure — and rewrite it as a polished prompt that can be pasted into a fresh AI agent session.

Principles

• Self-contained — the next session has zero memory. Write so an agent picking this up in a fresh window can follow it without re-asking the user.
• Faithful — never invent requirements, technologies, or assumptions the user didn't state. If something important is missing, see the clarification section below.
• Well-shaped — pick whatever shape serves the draft — a one-line ask, a paragraph, or sections — whichever a fresh agent will follow most easily. No template, no padding, no empty headings.
• Tone-neutral — drop filler ("please could you maybe…"), keep substance. Fix obvious typos silently as part of the rewrite.

Output format

A single block of plain markdown delivered directly in chat. Headings and lists inside the prompt are fine where they aid clarity.

No preamble, no chat framing, no closing remark. No "Sure, here is…", no "Hope this helps." The response IS the deliverable — anything wrapped around it is fluff.

Ask for context and clarification

Stop and ask the user when the draft is missing required context, internally contradictory, or ambiguous enough that you'd have to guess what they meant. A typo that creates ambiguity (e.g. a misspelled library name that could be one of two packages) is a clarification trigger, not a silent fix. Otherwise rewrite without asking.

When you do ask, list the missing pieces as a short bullet list and wait for the user to fill in before generating the refined version.