prompt-enhance

Prompt Enhance

Return one refined prompt a coding or knowledge agent can execute. Do not interview. When information is missing, use an assumption or [NEEDS: ...].

Workflow

1. Read the input and referenced context: repo, files, AGENTS.md, CLAUDE.md, README.md, relevant code.
2. Resolve unambiguous references such as "the thing" or "that script".
3. Identify defects: vague goal, buried intent, contradictions, missing scope, missing output format, missing success criteria, negative phrasing, premature-stop scaffolding, irrelevant noise.
4. Rewrite the prompt. Preserve intent. Remove filler. Keep substance.
5. Return the refined prompt, then Changed, Assumptions, and Gaps lines as needed.
6. Offer one action: run it now, save it to prompts/<slug>.md, or stop.

Principles

• Lead with the goal in one sentence.
• State output, format, constraints, and scope directly.
• Resolve contradictions or state priority.
• Say what is in scope and what is out of scope.
• Use numbered steps only when order matters.
• Prefer positive instructions.
• Include only context that changes the work.
• Specify output contract: format, structure, length, path, naming, overwrite behavior.
• State success criteria: tests, files, observed behavior, or other proof.
• Tell the agent to proceed on reasonable assumptions unless genuinely blocked.
• Use a role only when it changes tone or judgment.
• Separate instructions, context, examples, and pasted input when they could be confused.
• Add examples only for exact format or style.

Gaps

• Use a sensible default when safe. Record it under Assumptions.
• Use [NEEDS: <missing detail>] when no safe default exists.
• Do not ask questions. Use clarify for interviews.

Output

Refined prompt:

<rewritten, self-contained prompt that reads cold>

---
Changed: <substantive improvements>
Assumptions: <defaults baked in, omit if none>
Gaps: <remaining [NEEDS: ...] items, omit if none>

Then offer: run it now, save it to prompts/<slug>.md, or stop.

Example

Input:

script that pulls podcast mp3s and transcribes them, use whisper, don't make it slow, cache stuff somewhere sensible

Output:

Refined prompt:

Build a Python CLI script that downloads podcast episodes from an RSS feed and transcribes each MP3 with OpenAI Whisper.

Input: RSS feed URL as a required argument.
Output: save each MP3 and transcript (`.txt`, same basename) under `data/<podcast-slug>/`.
Caching: skip any episode whose MP3 and transcript already exist.
Performance: download and transcribe episodes concurrently.
Failure: exit non-zero for a bad feed URL; skip and log individual episode failures.
Success: running twice performs no redundant downloads or transcriptions.
Out of scope: UI, speaker diarization, non-MP3 formats.

---
Changed: added input, output, caching, concurrency, failure handling, and success criteria.
Assumptions: RSS feed input; local `data/` output.