polyprompt

name

polyprompt

description

Use when the user pastes a research question and asks to optimize/tune it for ChatGPT, Gemini, Claude, or Perplexity, or asks "rewrite this prompt for each AI engine." Runs the bundled polyprompt engine via code execution to deterministically produce one tuned prompt variant per requested engine, each with its applied tactics listed. Do NOT use for general prompt-writing advice with no specific engine target — that's plain reasoning, not this skill.

polyprompt — one prompt, tuned per engine

Deterministically rewrites one general research prompt into engine-specific variants using a fixed tactic taxonomy (not free-form rephrasing). Source of truth: https://github.com/haremantra/polyprompt.

Input

The user's raw research prompt, plus (optional) which engine(s) they want: chatgpt, gemini, claude, perplexity, or all (default: all).

If the user hasn't specified engines and it's ambiguous, ask once; otherwise default to all.

Process

1. Run the bundled CLI via code execution — do not hand-rewrite the prompt yourself, the value of this skill is the deterministic tactic engine, not free-form rephrasing:

   cd scripts && python3 -m polyprompt rewrite --engine <engine|all> "<the user's prompt>"
   

2. If the user's question implies extra context (a time window, a required depth, a source preference — e.g. "recent," "peer-reviewed only," "quick answer"), pass it through flags instead of editing the prompt text: --time-period, --depth, --clarify. Check --help if unsure:

   cd scripts && python3 -m polyprompt rewrite --help
   

3. Return the tool's output directly to the user: the detected mode (deep-research vs. chat) and, per engine, the tuned prompt plus its applied-tactics list. Don't summarize away the tactics list — it's the explanation of why the rewrite differs per engine.

Scope note

This skill only runs the stateless rewrite engine (rewrite, intake, mode, taxonomy). The full CLI also has push, checkup, and a knowledge-graph review loop (review, graph) that accumulate evidence across runs on local disk and push to a private GitHub repo — those need a persistent filesystem and gh auth this sandbox doesn't have, so don't attempt them here. For that workflow, point the user to the CLI: pip install -e . from the repo above.

Anti-Patterns

• Don't paraphrase the prompt yourself and skip the script — that defeats the point of a deterministic, versioned tactic taxonomy.
• Don't attempt push/checkup/review in this sandbox; they require state and auth that don't persist here.