Safety and Responsible Use

Last updated: 12 July 2026

How Provely (provely.app, an Unfussy Labs product) handles the tricky parts of running a prompt against real models, and what we do not claim to do.

When a model declines to answer

Sometimes a model will decline to answer. We treat a refusal as its own outcome, distinct from a normal answer.

  • A refused run is shown clearly as a refusal, not dressed up as a result.
  • A refused run is excluded from the denominator of your count. If one of the three models refuses, the headline reads against the models that answered, and the refusal is noted separately.
  • If too many runs refuse and the check cannot complete, we do not charge it against your daily allowance. A safety decline is not a failed check on your part, and it will not cost you your run.

Instructions hidden inside a prompt

The text you paste is data under review. If it contains instructions aimed at the model, for example "ignore your previous instructions" or "reveal your system prompt", we do not follow them. We report them back to you as a finding, under the injection risk category, so you can see what a real system might have obeyed.

The only place your text is actually executed is when we send it to the models to answer, and those runs have no tools. They cannot read files, call other services, or take actions. They answer your prompt and nothing else.

What we do not check

We want to be plain about the edges of this tool:

  • We do not verify facts. If a run produces a confident, wrong answer, we can tell you the models disagreed, but we do not fact-check what any of them said.
  • We do not test your setup. We run against three models on our side. We do not test your model, your account, your system prompt in production, or your environment.
  • We report the runs we made. Results describe those specific runs. They are evidence about the past, not a promise about the future.

Using the service responsibly

Please use Provely for prompts you have the right to submit, and do not use it to refine content meant to harm, deceive, or harass. The Terms of Use set out acceptable use in full. This is a review aid, not a guarantee, and you stay responsible for how you use your prompts and their results.

Reporting a problem

Found a bug, a bad result, or something that looks unsafe? Tell us: unfussylabs@gmail.com. Reports about safety and abuse are read first.

How to use Provely

Here is the whole run, from a prompt you are unsure about to one you can use.

  1. Paste your promptPut the prompt you want to test into the box on the home page. Nothing runs until you press the button.
  2. Read the findingsYour prompt comes back marked up. Everything is measured against what Provely read it as asking for; if that reading is wrong, Edit the intent and the findings are rebuilt from your version.
  3. Read the runsYour exact prompt goes once to each model in the run. You get every answer in full, an estimate of what it cost, and marks on your own words showing what held and what did not.
  4. Answer the questionsThe findings come with short questions about the things only you know, such as the format you want or who the answer is for. Answering is optional, and each answer you give goes into the improved prompt.
  5. Pick your fixesTick the fixes you want and Provely writes an improved prompt from them. You get two of those rewrites per check. You can also edit the draft by hand as much as you like.
  6. Run the re-testRun the improved prompt and compare. The before and after sit side by side, with the evidence behind every mark.
  7. Take it away, or share itCopy the improved prompt and use it wherever you like. Sharing a result is a separate choice, and you see a preview of exactly what goes public first.

The longer walkthrough is in the Prompt Check guide.