Skip to content

Text-Stil-Editor

The Text-Stil-Editor is a web-based application for the targeted transformation of texts into different language styles. Processing follows a two-stage LLM workflow with an optional neutralization preliminary step, followed by stylization driven by fine-grained, configurable controls. The application runs against a locally operated LLM instance; input texts do not leave the operator's own infrastructure.

At a glance

  • Convert texts into a required language style — for example for an academic paper, an official letter, or an application.
  • Neutralize existing texts before stylization to remove irony, emotional tone, or colloquial register as a baseline.
  • Adapt the language level to a specific audience (CEFR A1–C2, children to seniors).
  • Experiment with fine-grained controls to build and progressively refine a custom combined style.
  • Generate creative text forms — poem, fairy tale, satire, horror story, or advertising copy.
  • Compare several stylistic variants of the same text side by side.
  • Export individual control combinations as a compact code, share them, and reload them later.

Highlights

Compared to a direct query to an LLM or a simple prompt script, the Text-Stil-Editor separates stylistic transformation into clearly defined steps with explicit parameter control. This makes results more reproducible, comparable, and controllable in detail.

  • Two-stage workflow. An optional neutralization stage removes tone, register, redundancy, and other dimensions before stylization. This unifies the input baseline and yields more consistent stylization output.
  • Configurable neutralization. Ten separately selectable neutralization dimensions — seven safe and three experimental — allow controlled preprocessing that goes well beyond a generic simplification step. Neutralization can also be used as a standalone task without subsequent stylization.
  • 34 controls in three types. Polar controls (e.g. Formal ↔ Informal), intensity controls (e.g. irony 0–10), and stepped controls (e.g. language level A1–C2) cover tone, formality, clarity, creativity, rhetoric, format, and target audience.
  • Explicit intensity scale. Numeric control values are translated into a five-tier semantic (slight, moderate, distinct, strong, EXTREME) before being inserted into the prompt. The LLM thus receives unambiguous instructions instead of vague adjectives.
  • 23 presets. Predefined control combinations for common styles (e.g. academic paper, official letter, children's book, fairy tale, satire) as a starting point and as a reference for custom adjustments.
  • Configuration hashes. Any control combination can be exported as a compact, copyable code and reloaded from the same code. Settings can thereby be documented, shared, or attached to a publication.
  • Connection to 1 source. OpenAI-compatible API of a locally operated LLM instance (vLLM). Inputs do not leave the operator's own infrastructure.
  • Traceable results. Input text, neutralization intermediate, output text, and the full control configuration are stored per run in a session history and can be compared side by side.
  • Quantified input sizes. Token counting via tiktoken with limit display before the LLM call, allowing requests outside the configured range to be detected early.