Skip to content

TextWerkstatt

TextWerkstatt is a web-based application for drafting, revising, and structuring texts in the higher-education context. It combines agentic processing, a two-stage quality check with embedding-based fact verification, and domain-specific task profiles in a single interface. Tasks of varying scope are handled through two modes — a chat dialogue for shorter texts and a multi-stage pipeline for longer documents and complex undertakings.

At a glance

  • Convert meeting notes into formal committee or team minutes — either as a results-only or a full-record protocol, with verbatim resolutions and voting outcomes.
  • Draft response letters to students and replies to sensitive complaints — with a tone appropriate for higher education and standardised wording for legally sensitive topics.
  • Compose press releases, university communications, and job advertisements, or revise existing texts in a targeted way.
  • Review and consistently structure IT and project documentation, concept papers, and submitted offers.
  • Analyse and refine system prompts and image prompts for AI applications.
  • Consolidate multiple source documents into a single text and verify it against the sources.
  • Iterate on generated texts — shorten, adjust tone, replace sections, or supplement from material — and step back to earlier versions when needed.

Highlights

Compared to a direct LLM prompt or a simple script, processing is split into multiple phases, steered by domain-specific profiles, and verified against the source material. This affects completeness, factual fidelity, and consistency of the output.

  • Agentic pipeline for complex tasks. Longer texts are decomposed into analysis, work plan, section-by-section processing, quality check, and homogenisation. Each phase runs with separate prompts and different models, so that planning, drafting, and verification have their own responsibilities.
  • Domain knowledge per task type. Twelve predefined task profiles (committee and team minutes, press release, response letter, complaint reply, university text, image prompt, system prompt, project documentation, offer and concept review, job advert, complex subject matter) each carry their own quality criteria, tone specifications, and clarification logic.
  • Two-stage hallucination check. The first stage is a self-check by the model against the raw material; the second is a semantic embedding-based comparison of extracted statements (numbers, dates, names, quotations) against an inventory of the source texts. Statements without a sufficient match are flagged as suspicious.
  • Dual-LLM strategy. Quality-critical steps (analysis, planning, verification) use a stronger primary model; section processing and small edits use a faster secondary model. Both models are independently configurable; family-specific reasoning toggles are set automatically.
  • Intent router for follow-up iterations. After the initial generation, further inputs are classified (e.g. "shorter", "more formal", "rework section 3", "correct the wording of the resolution"). Unambiguous instructions trigger the matching operation directly; ambiguous ones prompt a clarification.
  • Scope model for targeted edits. Changes can be limited to the entire document, a section, or a single paragraph. For large documents, surgical edits are performed so that only the affected region is regenerated.
  • Versioning with branching. Each substantive change creates a version entry with timestamp and description. Earlier states remain accessible; resuming work after stepping back marks the skipped versions as branched without deleting them.
  • Briefing dialogue before long runs. Before pipeline runs, open profile questions are identified; only those that cannot be answered from the uploaded material appear in the clarification dialogue.
  • Connection to internal services. Connection to four backend services: a primary LLM for quality-critical steps, an optional fast LLM for sections and edits, an embedder for inventory and statement vectors, and a reranker for re-ordering retrieval candidates.
  • Export with provenance footer. Word, Markdown, and text exports include a footer with timestamp, task type, models used, and token usage. This makes it traceable how a text was produced.