m1nga/prompt-distill

Prompt Distill

Extract the intent; do not redesign the request.

Default behavior

• Preserve the user's framing, priorities, examples, language, and level of directness.
• Remove transcription noise, repetition, abandoned thoughts, and contradictions the user clearly corrected.
• Surface the actual outcome, essential requirements, and meaningful constraints.
• Keep useful ambiguity visible instead of silently inventing details.
• Prefer a small edit over a comprehensive rewrite. If the original is already clear, say so or return it nearly unchanged.

Trust the current model to reason. Do not add personas, generic expertise claims, step-by-step reasoning commands, motivational language, or obvious quality instructions.

Choose the response mode

Direct task

If the skill was invoked but the user is actually asking the agent to do something rather than asking for a rewritten prompt, resolve minor input noise internally and perform the task. Do not make the user review an intermediate prompt.

Reusable prompt

When the user explicitly wants a prompt to reuse elsewhere, return a polished version in the language of their input (or the language they name). Use natural prose or the smallest structure that improves execution.

For a complex request, include only the sections that add value:

• Core objective — the result the user actually wants.
• Requirements — distinct deliverables or decisions.
• Context or constraints — only details that affect the result.
• Open point — only a genuinely blocking ambiguity.

Do not force headings for a simple request. Do not append a translation or summary unless requested.

Interpretation rules

• Treat a corrected thought as replacing the earlier version.
• Treat "or should I…?" alternatives (in Chinese: "还是…？") as a request to compare and make or support a decision.
• A quick either/or inside a distill request is handled inline; an open-ended dilemma the user wants to explore belongs to thinking-partner.
• Treat examples as evidence of preference unless the user says they are exhaustive requirements.
• Treat "make the call" (in Chinese: "你看着来") as permission to choose within the stated scope, not permission to expand the scope.
• Keep separate projects separate when one dictation contains unrelated tasks.
• Resolve obvious speech-to-text errors from context; preserve uncertain names and flag them only when they matter.
• Use available conversation or workspace context when the user refers to earlier work. Do not manufacture missing history.

Quality check

Before responding, verify:

1. The core objective survived unchanged.
2. No real requirement, example, or boundary was lost.
3. The result is shorter or clearer than the input — not merely more formatted.
4. Nothing was added just to make the prompt look professional.