Welcome to the Language Sculptor

Core idea

The Analyze and Sculpt tabs are built around the idea that the overall style or flavor of a text passage can be characterized by a number of more or less independent stylistic dimensions, each quantified on a scale from -1 to 1.

• Analyze uses AI to estimate the values of each element of the selected profile.
• Sculpt lets you adjust those values, using sliders, to revise text or even generate new text from your own prompt.

The goal is to explore how each model responds when you emphasize specific traits—what becomes interesting, what falls flat, and where the underlying framework starts to break. A simple way to begin is to set all sliders to their neutral (zero) positions. Then choose one slider, move it all the way to one extreme, and generate or revise text. Next, move that same slider to the opposite extreme and try again. Comparing the two results will help you see whether that slider has a meaningful and recognizable effect.

The Modernize tab is deliberately separate—an independent (and still experimental!) helper for simplifying archaic prose. It reuses a bit of shared infrastructure but does not follow the slider framework.

Important limitations

• Not production-ready. Results are inconsistent. Treat every output as an experiment, not ground truth or polished prose.
• Passage length matters. Some profiles (and individual sliders) require longer samples before it's reasonable to try to estimate a particular metric. Paragraph-size or smaller text samples are insufficient for any metric that purports to be related to the structure or flow of a narrative, for example.
• Text type matters. Some profiles make sense only for narrative fiction, for example, while others are optimized for academic and expositional writing. You can of course try any profile with any sample of plain text (.txt, .md, maybe .tex) but don't be surprised if the results are sometimes unimpressive.
• Generation vs. sculpting. Certain sliders are meaningful only when analyzing a passage, not when trying to coerce new text or revise existing prose. Expect some controls to behave like suggestion knobs rather than strict directives.
• Model variability. GPT, Claude, and Gemini interpret the same vector very differently. Some adhere closely; others drift or ignore inputs entirely, especially on niche profiles.

How to explore

• Start with the Analyze tab: load a sample, inspect the inferred sliders.
• Apply those sliders to the Sculpt tab and tweak one dimension at a time to feel how the prose responds.
• Try longer passages (200–400 words) when evaluating complex dimensions like discourse organization or foregrounding.
• Compare providers: run the same vector through multiple models and note where they converge or diverge.

Above all, treat this interface as a sandbox for curiosity. The most valuable findings come from paying attention to failure cases as much as successes.