· 4 min read · AI, UX, Tools

Tweet style sliders

A hand moving a fader on a digital mixing desk
Photo: hanmaili, CC0

I built a small app that turns any text into a tweet, then lets you change the tone with eight sliders instead of a new prompt. People liked the control. They also told me which sliders blur together and what is missing: an undo button.

What it does

You paste any text into one big field. The app summarizes it into a tweet. Below it are eight sliders, each from 0% to 100%: formality, complexity, emotional intensity, seriousness, temporal style (traditional to contemporary), colloquialness, clarity and credibility. Move one and the tweet preview updates. You can try the app or read the project page.

I started with OpenAI's API, then switched to Mistral, which is free to use. The quality dropped a bit, so I added a second step that cleans up Mistral's output.

Why sliders and not a prompt

Jakob Nielsen wrote in 2023 that writing prompts is a usability barrier. Many people can't say in words what they want. His guess was that good AI interfaces would be hybrids, mixing typed intent with graphical controls that "show people what can be done". The Nielsen Norman Group made a similar point in August 2024: controls around the text field help people discover what a bot can do and lower the effort. Their examples are buttons, toggles, menus and cards. Sliders aren't in the list. I haven't seen a design guide for slider controls on AI output from before this date.

Sliders also fit an old idea. In 1983 Ben Shneiderman described direct manipulation: you act on the thing itself, and the actions are quick, small and reversible. Dragging a slider and watching the tweet change is that idea applied to text.

Big companies were moving the same way. Apple's Writing Tools let you rewrite text as professional, concise or friendly, as a few fixed choices. Anthropic showed the continuous version inside the model: researchers could turn a feature up or down, and Claude's behavior followed. A related research line, representation engineering, looks at controlling model behavior this way. My app does nothing so deep. It turns each slider value into instructions for the model. But it's the same wish: a dial instead of a sentence.

Where it went wrong

Users said some sliders overlap. That's no surprise in hindsight. What is the difference between formality and colloquialness, or between clarity and complexity? Eight dimensions of "tone" are not eight independent things. The Nielsen Norman Group's slider rules say sliders suit approximate values, and that a wide or finely divided range makes precise selection hard. Mine run 0-100 and there are eight of them. It's a lot to ask of a user.

The undo request is also fair. The classic heuristic is user control and freedom: people act by mistake and need a clear way out. A slider that rewrites the tweet on every move makes that worse. The earlier version is lost. My guess is that a model won't give the same tweet twice for the same values, so you can't just drag back. A history of previous tweets would fix it.

Models and cost

My switch from OpenAI to Mistral was about cost. Mistral announced a free tier on September 17, 2024, meant for trying and prototyping, and TechCrunch noted that production use may need a paid plan. That fits a hobby project, not a product.

OpenAI wasn't expensive either. GPT-4o mini, launched in July 2024, costs $0.15 per million input tokens and $0.60 per million output tokens. I can't say which OpenAI model suits this app best, but a short text and a tweet-sized answer would cost a tiny fraction of a cent. The real cost of live sliders is the number of calls. Every drag is a request.

The quality gap matches the research. A June 2024 study of text style transfer found that prompted models do reasonably well in English, that results differ a lot between models, and that dedicated training helps more than prompting.

My view

Sliders are a good first draft of a style control. I'd keep three or four, with clear labels at both ends, and drop the ones people can't tell apart. Credibility is the one I'm least sure about. A slider that makes a claim sound authoritative changes how the text sounds, not whether it is true. I'd think hard before leaving it in.

Sources

Comments