Prompt Tester

A prompt that “feels” better isn’t always structurally better — and eyeballing two drafts side by side rarely tells you which one will actually perform differently with an AI model. The Prompt Tester runs two prompt drafts against the same eight-point checklist and scores each one, so instead of guessing which version is stronger, you can see exactly which structural elements one has that the other is missing.

How to use it

1. Paste both drafts

   Your first version into Prompt A, your revised version into Prompt B.

2. Compare

   Click Compare Prompts to run both through the same eight-point checklist.

3. Read the summary

   Check each scorecard, and the summary line for the biggest gap between the two.

Why it helps

The checklist looks for the same things a careful prompt writer checks by hand: a defined role, a stated task, an output format, context, constraints, examples, and language specific enough to avoid vague filler words like “some,” “good,” or “things.” Each of these maps to a concrete gap in how a model interprets an instruction — a missing format means the model guesses the shape of the answer, a missing constraint means it guesses the length, and so on.

What the tester deliberately doesn’t do is judge writing quality. Two prompts can score the same on structure and still read very differently — one clear and direct, one clunky but technically complete. That’s on purpose. Structure and style are separate problems, and conflating them is how people end up “fixing” a prompt’s wording when the actual issue was a missing constraint the whole time.

What the checklist actually checks

Each of the eight checks targets a specific, common failure mode in prompt writing:

• Role or persona — does the prompt tell the model who to be before it tells it what to do?
• Clear task up front — does it open with a concrete action verb (write, summarize, compare) rather than background information?
• Output format — is there any indication of what shape the answer should take?
• Context or background — is there enough surrounding information for the model to calibrate its answer, or is it working blind?
• Length or constraints — is there a stated limit, whether that’s a word count, a rule to follow, or a hard requirement?
• Examples — does the prompt show what “done” looks like, rather than just describing it?
• Specific rather than vague language — how many filler words like “some,” “nice,” or “stuff” made it into the final draft?
• Reasonable length — is the prompt long enough to actually carry the information above, or is it too thin to do any of this?

A typical before-and-after

Someone testing “write something about our new product” against a revised version — “You are a product marketer. Write a 100-word announcement for our new running shoe, aimed at existing customers, in an upbeat but not salesy tone” — will usually see the second draft clear five or six checks the first draft failed entirely. The score isn’t the point; the list of which specific checks flipped from failing to passing is, because that list is a direct to-do list for improving the weaker draft.

Who this is for

Anyone iterating on a prompt they’ll reuse regularly benefits from this — content teams standardizing a prompt template across a company, developers refining a system prompt before it goes into production, or anyone who’s rewritten the same prompt five times and wants an objective way to tell if version five is actually better than version two. It’s also useful paired with the Prompt Optimizer: test a draft here first, then run it through the optimizer to generate a scaffold for whatever’s still missing.

Tips for better results

FAQ