Prompt Review Scoring and Improvement

Problem

Prompt review often struggles with two issues: session history mixes small talk, repeated confirmations, and fragmented follow-up questions, so direct scoring can be misleading; and a single turn often lacks context, constraints, and output format, making feedback feel vague. This skill narrows the task to auditable, reusable prompt evaluation.

How It Works

It first applies noise filtering, keeping only turns with substantive instructions, requirement statements, or condition-setting. Then it performs context aggregation, merging 2-5 consecutive turns that refine the same goal into one prompt evolution case. It scores each case on three dimensions: intent clarity, context completeness, and constraint explicitness, averaging them into a 1-10 score. For low-scoring cases, it provides a concrete rewrite template with [variable] placeholders and three immediately actionable improvement suggestions.

Notes

The skill emphasizes honest evaluation and does not inflate scores to please the user; a vague request is flagged directly. It also writes detailed results to a Markdown file while keeping terminal feedback minimal, so it fits daily prompt audits and template accumulation rather than long inline analysis.