Skip to content

TextWerkstatt — Features

TextWerkstatt covers typical writing and revision tasks of the higher-education context. The feature set spans initial generation of formal documents from raw material, iterative detail edits, and export in multiple formats with a provenance footer.

Use cases

  • Committee and team minutes from meeting notes. Meeting notes are converted into a formal protocol — as either a results-only or a full-record version, with resolutions, voting outcomes, attribution by name, and a complete formal header.
  • Response letters and complaint handling. Replies to student inquiries and to sensitive complaints are drafted with a tone appropriate for higher education; legally sensitive topics are flagged accordingly.
  • Press releases and university communications. Background materials, key facts, or quotations are turned into texts suitable for outward-facing communications — optionally in several length variants.
  • System prompts and image prompts. Existing prompts for AI applications are analysed, missing dimensions are identified, and revised, clearly structured versions are produced. Image prompts for teaching materials are formulated according to target medium and visual style.
  • Documentation and concept work. Project documentation, IT descriptions, concept papers, and incoming offers are checked for consistency, completeness, and requirement coverage, and restructured where appropriate.
  • General text work with multiple source documents. Several uploaded files (e.g. notes, reports, mail threads, tables) can be consolidated into a single text, verified against the sources, and revised together.

At a glance

  • Two processing modes: chat dialogue for shorter texts, multi-stage pipeline for complex tasks.
  • Twelve task profiles with their own quality criteria, plus a free-form task without profile binding.
  • File import for seven formats (txt, md, csv, html, docx, pdf, pptx) up to 50 MB per file.
  • Export to Word (.docx), Markdown (.md), and text (.txt) — each with a provenance footer.
  • Two-stage hallucination check with embedding-based fact verification against the source material.
  • Version management with branching; up to 20 prior states accessible.
  • Intent-based follow-up iterations with scope on the entire document, a section, or a paragraph.

Task profiles

The application includes twelve predefined task types, organised into three categories.

Create — committee meeting minutes, team meeting minutes, response letters to students, replies to sensitive complaints, press release, university text, image prompt for teaching materials.

Improve — system prompt, IT and project documentation, offer and concept review, job advertisement.

Both — preparation of complex subject matter.

Each profile contains domain knowledge, tone specifications, a set of quality criteria, and typical clarifying questions that are raised before processing begins, unless they can be answered from the material. An additional "free task" option processes content without a fixed profile.

Processing modes

Quick mode. Chat-based initial generation for shorter tasks. The input is processed directly by the primary model with the profile-specific system prompt; iterations occur within the same session.

Workshop mode. Multi-stage pipeline for complex tasks or extensive material. The phases are: a survey of the material, creation of a work plan with section structure and quality criteria, section-by-section processing, automatic quality verification with an optional correction loop, and final homogenisation of the entire document. The workshop mode is selected automatically when the material exceeds a configurable threshold or when the task type is marked as inherently complex; manual selection is also supported.

Connectors and data sources

TextWerkstatt uses four backend services and a file upload interface.

  • Primary LLM (internal). Internal, OpenAI-compatible model used for quality-critical steps: analysis, planning, quality verification, correction, and homogenisation.
  • Fast LLM (internal, optional). Internal, OpenAI-compatible model used for section-by-section processing and small, fast edits. Falls back to the primary LLM automatically if not configured.
  • Embedder (internal). Internal service for vector embeddings. Used both for inventorying source material and for the semantic hallucination check. Supports several OpenAI- and HuggingFace-compatible endpoint variants and detects the appropriate scheme on its own.
  • Reranker (internal). Internal service for re-ordering retrieval candidates. Refines the order of entries proposed by the embedder before they are passed on to the pipeline.

File import

Uploaded files are converted into a uniform Markdown text before processing. Supported formats are plain text (.txt), Markdown (.md), CSV tables (.csv), HTML (.html with script and style filtering), Word documents (.docx with preservation of headings, lists, tables, block quotes, and inline formatting), PDF (.pdf via text extraction), and PowerPoint presentations (.pptx on a best-effort basis). A whitelist of permitted file extensions and a size limit guard the import against unsuitable content.

Export formats

  • Word (.docx). Structured Word export with heading levels, lists, tables, code blocks, block quotes, and inline formatting. Font, page margins, and spacing are preset.
  • Markdown (.md). Markdown export with the provenance footer as a block at the end of the document.
  • Text (.txt). Plaintext export with Markdown markup removed and a provenance footer as an ASCII block.

Each export contains a provenance footer with creation timestamp, task type, model names used, and cumulative input and output tokens.

Quality assurance

  • Domain-specific quality criteria. Each task profile carries a set of criteria (e.g. verbatim adoption of resolutions for minutes, completeness of formal header data, requirement coverage for concept reviews). These criteria are checked against the generated document during automatic quality verification.
  • Hallucination check via the model. Quality verification includes an explicit comparison of the document against the raw material. Names, numbers, dates, and facts not supported by the material are reported as a defect of type "hallucinated".
  • Semantic fact verification. When an inventory of the source material exists, an additional embedding-based second check is performed. Statements such as numbers with units, dates, person names, and verbatim quotations are extracted, embedded, and compared against the inventory. Statements without sufficient similarity to an inventory entry are flagged as suspicious.
  • Correction loop. When defects are reported, a correction pass runs through the primary model followed by re-verification; the number of passes is bounded to keep processing time manageable.
  • Briefing before processing. Profiles with a configured briefing check which of the typical clarifying questions are already answered by the uploaded material and only ask the open ones.
  • Visible work plan. Before processing, the generated work plan is shown — including section structure, key elements, identified gaps, and global quality criteria.
  • Version management. Substantive changes create version entries. Earlier states remain accessible; resuming work after stepping back marks the skipped versions as branched, rather than discarding them.
  • Provenance footer. Each export carries a timestamp, task type, model names, and token usage — a basis for traceability and reproducibility.

Iterative editing

After the initial generation, follow-up inputs are classified by an intent router. Unambiguous instructions such as "shorter", "more formal", or "remove section 3" are executed without clarification; ambiguous inputs trigger a targeted clarifying question. A scope manager limits the operation to the entire document, a section, or a single paragraph, so that edits to large documents are made selectively.