AI prompt generator and builder

It shows the structure. A typical generator hands back a finished string and never explains what changed or why. Here you see which parts a prompt consists of, which you filled in, which are missing and how much each affects the result. Someone else’s prompt works exactly the same way: paste it into the same field and the breakdown appears on its own — there is no separate place for it.

No, and that is deliberate: a prompt tuned to one model's quirks stops working the week that model is updated. What the score measures instead is whether the prompt carries the parts that any model needs to be told — the task, the role it should adopt, the audience, the format of the answer, the constraints, the examples. Those are properties of the request rather than of the system reading it, and a prompt that has all of them works better on every model than a prompt missing three of them works on the one it was written for. Where a model genuinely differs is in length limits and in whether it accepts images, and those are practical facts rather than phrasing tricks.

From Chrome 148 and recent Edge the browser carries a language model inside it: it downloads once, lives on disk and computes on your device. The tool probes for it and writes on the button itself which path a run will take. Firefox and Safari have no such model yet — there improve and test go through our neural network, while building, scoring and the library work the same everywhere.

Because English is the native language of those models and they follow instructions in it more precisely. We measured this on our own music generator: a "no drums" instruction written in Russian was ignored in all six runs, while the English wording was obeyed in all six. So strengthening an image, video or music prompt returns English, while chat and code prompts come back in the language you wrote.

From the composition. Each kind of prompt has its own set of meaningful parts with its own weights: for an image, lighting weighs twice as much as palette, because lighting redefines the whole frame while palette only tints it. The score adds up the parts found, subtracts empty words, and cannot go above 40 without the main field.

In the browser, on a model that recent versions of Chrome and Edge carry inside them — the model ships with the browser, so there is no key to paste and no account to open. That model is small and it is not the point: the value is in the loop rather than in the answer. Change one part of a prompt, run it, see what moved; add the missing audience, run it again, see the shape of the reply change. That is how you learn which parts of a request actually do work, and it costs nothing to repeat twenty times. What you should not do is judge a prompt's final quality here — a small local model and a large hosted one answer differently, and the prompt is being tested for structure, not for the reply it produced.

The prompts you build are kept in your browser's localStorage, so they survive a reload and you can come back to them instead of rebuilding. The button in the "Your prompts" section clears them.

The prompts themselves are just text — instructions for a model — and we place no restrictions on using them. What you may do with the RESULT depends on the terms of the model you run it on: Midjourney, Suno, ChatGPT and the rest each set their own rules for commercial use.

Because Suno and Udio have two fields with different grammars. Style takes the description of the sound — genre, mood, instruments, tempo and key, vocal, exclusions — as one comma-separated line (up to 1000 characters on Suno). Song structure does not go there: it goes into the Lyrics field as [Intro], [Verse 1], [Chorus] tags above the lines. We tested on live Suno what actually matters: with a Russian-language style the model hit the requested key in one clip out of twelve, with English in 24 of 28, and it ignored a Russian “no drums” in all four clips. The separator makes no difference — commas and vertical bars “|” gave the same result — so the style is assembled with commas, as both models document.