Skip to content

Features

The feature set is organized into four areas: text transformation itself, management of the style controls, separate configuration of neutralization, and a history for tracing and comparing results.

Use cases

  • Writing an academic paper. An existing draft is reshaped using the "Scientific" or "Academic Paper" preset, applying hedging, precision, factual tone, and passive constructions. The "Language Level" stepped control allows fixing the output at C1/C2.
  • Official correspondence and applications. Texts are transferred into a formal, polite, factual register. Presets such as "Official Letter", "Application Letter", or "Business Email" provide an initial configuration that can be refined afterwards.
  • Language level adaptation. Specialist texts are simplified for a particular audience — for instance to CEFR A1/A2 for language learners or to a children's level via the "Children's Book" preset. The "Language Level" and "Target Audience Age" controls can be combined.
  • Creative writing. Texts can be transformed into poems, fairy tales, horror stories, satires, or advertising copy. Intensity controls (Poetic, Magical, Ironic, Sarcastic, Dramatic) allow fine-grained dosing of individual stylistic devices.
  • Neutralization as a standalone task. Texts with pronounced tone, ironic content, personal evaluations, or rhetorical devices can be processed via neutralization only, without subsequent stylization. The result is a factual, neutral text that can serve as a starting point for further work.
  • Reproducible replication of a language style. A combination of multiple active controls together with a stored control configuration enables a desired style to be rebuilt reproducibly and applied to additional texts.

At a glance

  • 34 style controls in 7 groups, 3 control types (polar, intensity, stepped)
  • 23 predefined presets, alphabetically sorted
  • 10 neutralization dimensions (7 safe, 3 experimental)
  • Optional two-stage workflow: neutralization → stylization
  • Saving and loading complete control configurations via copyable codes
  • History with up to 20 entries, side-by-side comparison of two versions
  • JSON export and import of control definitions
  • Token counting via tiktoken, live display, and limit check

Text transformation

Transformation can be used in two modes: single-stage as stylization only, or two-stage with prior neutralization. With neutralization enabled, the text is first cleaned of the selected dimensions and the intermediate result is used as input to the subsequent stylization. Both stages remain inspectable.

  • Control configuration in the UI. Controls are presented grouped in collapsible accordions, can be activated individually, and are set via sliders to values between -10 and +10 (polar) or 0 and 10 (intensity). Stepped controls use a selection list.
  • Presets. A dropdown selection sets the corresponding controls in a single step. Manual adjustments afterwards remain possible at any time.
  • Saving and loading via code. Active control configurations are presented as a compact, base64-encoded code that can be copied, stored, or shared. The same code restores the state exactly.

Neutralization

Neutralization is configurable in a separate tab. It comprises ten independently switchable dimensions:

  • Safe dimensions. Tone, register, redundancy (active by default), as well as word choice, subjectivity, perspective, and intensifiers.
  • Experimental dimensions. Structure (e.g. rhetorical questions → statements), metaphor (figurative language → literal description), and cultural references. These intervene more deeply in the text and are flagged as experimental.

Neutralization can be operated standalone — without subsequent stylization — and then yields a factual, neutralized text as the final result.

Style control management

A separate tab allows controls to be added, edited, and deleted. Each control is defined by ID, name, type, description, and — depending on type — opposite or list of steps. The full set of controls can be exported and imported as JSON, supporting transfer between instances or version control in a repository.

History and comparison

Each transformation is stored with input text, intermediate, output text, selected control configuration, and, where applicable, the preset name in a session history (default limit: 20 entries). In the "History" tab, two entries can be selected and compared side by side — both their output texts and the underlying settings.

Connectors

The application accesses one external component:

  • Locally operated LLM instance (vLLM). Connection via the OpenAI-compatible chat completions API. Model name, base URL, and API key are configured via environment variables. The application makes no assumptions about the model in use and does not rely on model-specific features beyond the standard interface.

No further external data sources are integrated. Input texts and control configurations remain within the session and are not persisted.

Import and export formats

  • JSON. Complete control definitions (schema of default_regler.json) can be exported and imported.
  • Configuration code (base64). Active control settings of a transformation can be exported as a copyable string and re-imported.
  • Markdown output. The transformed text is rendered as Markdown in the UI, which makes structured results (headings, lists) visible.

Quality assurance

  • Two-stage workflow. The separation of neutralization and stylization reduces the overlap of conflicting instructions in a single prompt and yields more consistent results.
  • Explicit intensity semantic. Numeric control values are translated server-side into a five-tier scale (slight, moderate, distinct, strong, EXTREME) with a clear textual instruction.
  • Strict prompt rules. The system prompts suppress preambles, meta comments, and task repetitions; they require the core meaning of the text to be preserved.
  • Retry logic. Failed LLM requests are retried with a configurable retry count and wait time; the per-request timeout is configurable.
  • Token validation. Inputs are checked against a configured token limit before the LLM call; a colored status indicator (green/yellow/red) signals utilization.
  • Reproducibility. The exportable configuration codes and the history allow transformations to be re-run with identical settings and compared.