Guides

Prompt Check

Prompt Check shows you what a prompt actually does in one pass, with evidence rather than an opinion. You paste your prompt, it reads your intent back to you, then it runs your exact wording once on each of three different models and shows what each one did.

Nothing runs until you press the button, and the first thing you see is what Provely thinks you are asking for. That is your chance to correct it before any model spends a token.

What you get

In one check you go from a prompt you are unsure about to a clear picture of how it behaves, and a fixed version you can take away.

Your prompt marked up straight away: what Provely read it as asking for, and where it is vague or missing something.
Your exact prompt run once on each of three models: Claude Sonnet 5, OpenAI GPT-5.6 Luna and Google Gemini 3.6 Flash.
Each answer shown in full, with its token count and an estimated cost.
A marked-up view of what held and what did not, with the evidence behind every mark.
A verdict on which model did what you asked, and the option to apply fixes and run again.

Before anything runs

Provely works out what your prompt is asking for and marks it up on that basis, with no pause. If its reading is wrong, press Edit the intent on the findings screen, put it right in one sentence, and the marks are rebuilt from your meaning rather than its first guess.

A prompt can be up to 10,000 tokens, which is a long brief or a system prompt with examples. If yours is longer, Provely tells you the count and asks you to trim it before anything runs.

The three runs

Your prompt goes to Claude Sonnet 5, OpenAI GPT-5.6 Luna and Google Gemini 3.6 Flash, word for word, with nothing added. Each model answers, and Provely shows the answer, the tokens it used and an estimate of what that run cost.

The headline is a count, never a score out of a hundred. You see something like "2 of 3 models did what you asked", because that is a fact about the runs in front of you, not a rating we invented.

How the marks are earned

Marks are coloured by evidence, and they have to be earned:

Green marks an instruction that held across the runs, shown with the lines that prove it.
Orange marks a suspicion the rubric raised that the runs did not actually confirm, so you can weigh it yourself.
Red appears only after a clarify-and-rerun proves a real fault, never on a hunch.

That order matters. A worry is not painted red until the evidence is there, so the strong marks mean something when you see them.

Fix it and run again

When the findings point to a fix, you pick the ones you want and Provely rewrites your prompt to apply them. You get two rewrite attempts per check, which is enough to try a version, read it, and ask for one more take if the first is not quite right. You can also edit the draft by hand as much as you like; only the AI rewrites are counted.

Running your prompt as you wrote it and running it again with fixes still counts as one run for the day. When you are happy, you hand the improved prompt off to use wherever you like.

A worked example

Worked example

A support prompt that looks fine until you see how the models read one line.

Prompt: You are a support assistant for Acme’s billing team. Answer the customer’s billing question, summarise their invoice history first, and reply quickly.
Read-back: You want a friendly billing reply that recaps the customer’s invoices before answering, and you value speed.
What happened: "Reply quickly" pulled two of the three models towards short answers that skipped the invoice summary you asked for. The instruction to summarise held on only one model.
Fix: Drop "reply quickly" and make the order explicit: "First summarise the last three invoices, then answer the question." On the re-run, all three models kept the summary.

The count moved from 1 of 3 to 3 of 3, with the evidence for each mark shown beside it.

Daily allowances

Read-backs, full runs and rewrites each have their own daily allowance so the free service stays fair for everyone. See the usage limits for what resets and when.

Ready to try it on a prompt of your own?

Open Prompt Check