← the studio
A MoreSalamander · StudioLabs production

Prompt Polish Studio

A web app that scores your prompt across a five-dimension rubric, asks clarifying questions, rewrites it with your answers, and compares the outputs side by side — the methodology turned on prompt engineering itself.

Score the prompt. Then make it score.

What it is

A prompt isn't done until it clears the rubric.

Most prompt "improvement" is vibes. Prompt Polish Studio makes it checkable: a deterministic rubric scores the prompt across five dimensions, and a low score isn't a verdict — it's a worklist. The app asks exactly the questions the rubric flagged, folds your answers in, and re-scores. The studio's first publicly shipped project — a hackathon entry that proved a multi-step AI agent pipeline could ship under deadline.

Watch it run

A weak prompt gets scored, questioned, rewritten — and re-scored.
prompt polish · pipelineattempt 1
analyze
rubric score
gate ●
questions
polish
compare
rubric
— /100 · need ≥ 80
Five dimensions — clarity, specificity, context, constraints, output format. A low score names exactly what's missing, which becomes the clarifying questions. The rewrite folds your answers in and re-scores; the A/B compare shows the lift. The scorer is a rubric, not a vibe.

The three moves

Explain · Synthesize · Verify — applied to a prompt.
01 · Explain

The five-dimension rubric is the constraint, written down first — clarity, specificity, context, constraints, output format. It's the bar the prompt is measured against, before any rewriting happens.

02 · Synthesize

The model generates clarifying questions for whatever the rubric flagged, then rewrites the prompt incorporating your answers — turning a vague ask into a specific, well-formed one.

03 · Verify

The rubric re-scores the rewrite and the app runs an A/B compare of the outputs, side by side — so the improvement is shown, not asserted. Below the bar, it loops.

Best at — and the honest limit

A measurable lift you can see, not just trust.

Prompt Polish Studio is a React app that takes a rough prompt to a well-formed one and proves the gain with a side-by-side. It didn't win its hackathon, but it shipped — the studio's first public proof that a working multi-step AI agent pipeline could land under deadline.

What it can't do: the rubric scores a prompt's form — clarity, specificity, structure — not whether the prompt will get you the answer you wanted. It makes a prompt well-built; whether well-built is right for your goal is still your call, which is what the A/B compare is for.