Skip to content

Style

Style is a browser-based application for experimenting with text style transformations. It combines a structured, multi-dimensional schema for describing writing styles with a locally operated large language model. Style profiles can be derived from example texts, generated from a style description, or refined iteratively through before-and-after pairs. Profiles are portable and can be applied to new texts, visualized, and shared within a team.

At a glance

  • Derive style profiles from example texts and make their characteristics visible in a structured form.
  • Sharpen style profiles through repeated before-and-after feedback.
  • Convert texts experimentally into a different style and edit the result directly in the browser.
  • Save writing styles as JSON or ZIP packages and exchange them within a team.
  • Display style characteristics as a visual overview and use them for training purposes.
  • Get started without preparation using the bundled profile library.

Highlights

In contrast to a direct LLM prompt, Style works with an explicit, reusable style model. Style characteristics are not described ad hoc but captured in a structured form, persisted, and applied to texts in a controlled manner. This improves reproducibility and comparability of results.

  • Structured profile schema: Writing styles are captured along four dimensions — linguistics (sentence length, complexity, passive voice ratio, hedging), lexis (technical terminology, foreign words, abstraction level), pragmatics (formality, mode of address, target audience), and structure (paragraph length, headings, lists).
  • Three complementary paths to profile creation: Derivation from example texts, generation from a plain description of the target style, or iterative refinement through before-and-after pairs. The paths can be combined.
  • Iterative refinement: Up to five before-and-after pairs can be analyzed cumulatively to derive recurring change patterns from concrete edits.
  • Portable profiles: Profiles are exported as JSON or as a ZIP package (containing profile parameters, generated prompt, and example sentences) and can be stored externally, versioned, and exchanged.
  • Visualization of style characteristics: A dedicated overview displays profile parameters with text-based bars; suitable for training and for comparing profiles.
  • Locally operated LLM: The application connects to an LLM via an OpenAI-compatible API. In the intended configuration, the model runs locally (e.g., via LM Studio or Ollama) — texts therefore do not leave the local infrastructure.
  • Multiple input and output formats: Texts are imported as TXT, Markdown, or DOCX; results can be returned in the same formats. Structures such as headings and lists are preserved.
  • Bundled profile library: A collection of prepared profiles serves as a starting point for experiments — ranging from text genres (interview, bureaucratic, technical documentation, customer mail) to characteristic author styles.
  • Containerized operation: The application is delivered via Docker and prepared for operation behind a reverse proxy.