The Prompt Quality Score, and How to Raise It

Every optimization comes back with a quality score from 0 to 100. It is one number, not a stack of sub-scores. Read it right and it tells you whether a prompt is ready to ship or still leaking ambiguity. Read it wrong and you either over-polish a prompt that was already fine or trust one that quietly underperforms.

This guide covers what the number means, why a "good" prompt usually lands in the middle of the range, and how to use deep analysis to find the subtle issues a score alone won't name. Then you iterate.

What does the quality score actually measure?

The score is a composite. VantagePrompt judges your optimized output across several dimensions — clarity, specificity, structure, completeness, and linguistic quality — then weights and rolls them into one 0-100 figure. You see the total, not the breakdown. Clarity and specificity carry the most weight, so a prompt that is well-organized but vague about format, audience, or constraints will be capped by its weakest, heaviest dimensions.

The judge is deliberately strict. It is told not to inflate scores: most prompts are expected to average mid-rubric, and only genuinely exceptional prompts hit the top. That changes how you should read the band.

Why do two prompts for the same task score differently?

Specificity is where most points are won or lost. Compare these two prompts for the same task:

Write a summary of this quarterly report for the team.

Summarize this quarterly report for the engineering leads.
Output: 5 bullet points, each one sentence.
Focus on: shipped features, missed deadlines, and next-quarter risks.
Tone: factual, no marketing language. Skip financials.

The second names the audience, the output shape, the focus, and what to exclude. Those are exactly the things the specificity and completeness dimensions reward, which is why a tighter prompt lands higher even when both are grammatically clean.

What does deep analysis find that the score cannot?

The score tells you how good a prompt is. It does not tell you why it fell short. That is what deep analysis is for. It is an opt-in critic you trigger from the optimizer — a single pass that surfaces at most three subtle issues a simple linter would miss: semantic ambiguity, hidden assumptions, contradictory signals, scope drift, missing antecedents, or unclear success criteria.

Each issue comes back with a severity. A warning flags something that holds the prompt back; info is a nice-to-fix. Use the warnings as your edit list. A typical result looks like this:

{
  "issues": [
    {
      "severity": "warning",
      "label": "Ambiguous scope",
      "detail": "'The team' is undefined — engineering, sales, or all staff?"
    },
    {
      "severity": "info",
      "label": "No length bound",
      "detail": "Output length is unconstrained; add a sentence or word cap."
    }
  ]
}

What does the iteration loop look like?

Put the score and deep analysis together into a tight cycle:

1. Optimize and read the score. Low or mid? There's room to improve.
2. Run deep analysis to name the specific weaknesses.
3. Edit the raw prompt: resolve every warning first, then the info items worth fixing — usually adding format, audience, constraints, or examples.
4. Re-optimize and compare the new score against the old one.
5. Stop when the warnings are gone and the score plateaus. Don't grind for the last point.

For tasks that need step-by-step thinking, pair this with the reasoning use case and effort controls; for prompts that depend on current facts, grounding with web search closes the gaps deep analysis flags as hidden assumptions. The score keeps you honest; the loop is what moves it.