Features

LernWerkstatt provides five successive working steps as a wizard interface: learning profile, planning, release, production and result. Its scope covers the analysis of uploaded documents, goal-oriented topic planning, 18 didactic building blocks, a deterministic verification layer, and delivery as a single offline-capable file.

Use cases

• Entering a new subject area — From a question such as "Why do plants need water?" and the prior knowledge given, a unit emerges that answers exactly that question instead of reproducing a curriculum. The depth of treatment per concept follows the learning goal.
• Teaching material from existing documents — A standards document, a book chapter or lecture slides are uploaded and form the basis. Terminology comes from the documents, claims are checked against them, and every concept carries a provenance note.
• Self-study unit for exam preparation — A unit with high interaction density, flashcards for terminology, matching tasks for easily confused items and a final test. The file runs without a network, so it works in an exam room or on the move.
• Explaining quantitative relationships — Where a matter depends on parameters, a simulator with sliders or an interactive chart with bound controls is produced. The learner tries the relationship out instead of having it described.
• Multilingual material — The same learning intent can be produced in five languages. The language setting is an acceptance criterion rather than a prompt addendum; language breaks and foreign scripts are reported.
• Low-barrier learning material — The generated file is keyboard-operable and marked up for screen readers. Diagrams and graphics carry mandatory alternative texts, and formulas are set as MathML and thus readable aloud.

At a glance

• Five working steps with two checkpoints at which the plan can be corrected
• 18 building blocks across four reliability stages, from deterministic configuration to verified computational function
• Four treatment classes and two depth profiles governing scope and effort
• Deterministic verification layer without model calls, with automatic repair and degradation
• Five languages, with language discipline as an acceptance criterion
• Import: PDF, Word, PowerPoint, HTML, Markdown, CSV, plain text (50 MB per file)
• Export: single-file HTML, plus chapter scripts as Word document and Markdown
• Server-side production with fine-grained resumption and per-chapter rework

Building blocks

Eighteen block types are available, grouped by reliability stage. Which block goes where follows from the type of content: processes become diagrams, comparisons across several attributes become tables, parameter dependencies become explorable graphics, easily confused items become matching tasks.

Stage 1 — deterministic, schema-checked. Prose, callout and note boxes, tables, accordions, code blocks, quizzes with reasoned feedback, cloze exercises, matching tasks, flashcards, prediction tasks, plus open-ended exercises and error analyses in the application section.

Stage 2 — declarative grammars. Mermaid diagrams for processes and structures, charts via Chart.js or Vega-Lite for quantities and trends, formulas as MathML. Vega-Lite graphics can be interactive through bound parameters — sliders, select fields, linked views — without any code execution.

Stage 3 — verified computational function. Simulators consist of a pure function with declared parameters. Before delivery they are trial-run in a sandbox with a time limit.

Stage 4 — sandboxed widget. Complete HTML in a sandboxed iframe, only with explicit justification in the detailed plan.

Topic planning and depth of treatment

From prior knowledge, learning goal and source documents the application derives a concept inventory. Each concept receives a treatment class:

• V — derive in full, activate and practise
• K — cover compactly
• D — handle as a definition
• R — glossary entry only

The classification governs scope, model choice and exercise obligations. What counts is the question actually asked, not the subject's own sense of importance: what the person wanted to know becomes V; what they need for that, K; whatever else the field holds is dropped. An interference assessment supplements this — where new knowledge collides with existing knowledge, the derivation goes deeper. Two depth profiles (compact, detailed) govern the overall scope.

A concept graph then follows, with exactly one introduction site per concept, named prerequisites and points of reuse. It is checked for cycles and ordering violations before taking effect.

Analysis of uploaded documents

PDF, Word, PowerPoint, HTML, Markdown, CSV and plain text are read; archives are unpacked, with size limits guarding against decompression bombs. The extracted full text serves three purposes:

• It is the authority for the terminology used.
• It is the reference for fact checking — including table cells and the numeric values in charts.
• With an embedder configured, it is segmented into typed units (chapter, section, definition, example, procedure) and made searchable.

Where a reranker is also available, a coverage analysis maps supporting passages to each concept and conversely surfaces material that the plan does not mention. Concepts without support are marked as originating from model knowledge, and the text section is told to mark nothing as a quotation.

Terminology and glossary

A register separates established terms — from the concept inventory, the source documents, or approved after review — from newly appearing candidates. Only established terms become binding for later chapters; without that separation a coinage from chapter 2 would become a requirement for all the rest.

What is checked is whatever the text treats as a term: glossary entries, concept names, and terms introduced by a definition pattern. Candidates are filtered in stages — whitelist, source corpus, compound decomposition — and only the remainder is put to a model. Rejected terms land, with a replacement suggestion, on a list applied thereafter without any model call.

Within the unit, all glossary terms are discreetly marked in the running text. Hovering shows the short explanation, clicking leads to the full entry. The marking tolerates inflection and leaves code, formulas, headings and existing links untouched.

Quality assurance

Between model and delivery sits a verification layer working largely without model calls:

• Schema and block-contract validation for all 18 block types
• Quiz logic: exactly one correct answer in single choice, no identical options, no question without a wrong answer
• Matching tasks: uniqueness in both columns, no trivial pairs
• Cloze exercises: two-way placeholder consistency, warning where the answer is readable in the surrounding sentence
• Diagrams and charts: syntax probe, data consistency, unused parameters, missing alternative texts
• Interaction density and distribution, variety of forms across the unit
• Format and language discipline: Markdown and LaTeX residue, foreign scripts

Three model-supported checks complement this: a critic per lesson against didactic criteria, fact checking against the source documents, and a redundancy pass across the whole unit that finds repeated explanations.

Whatever can be repaired deterministically is repaired before checking. Whatever remains broken is replaced by its own description as text rather than blocking the unit; the report records the substitution.

The delivered learning unit

The result is a single HTML file with no external requests. Renderer libraries are embedded and formulas set as MathML at build time. It contains:

• Full-text search across the entire unit
• Direct links to individual sections
• A filterable glossary with a jump back to the introducing lesson
• Glossary tooltips at every occurrence in the text
• Print formatting
• Keyboard operation and ARIA markup

If the final gate is not passed, a file is produced nonetheless — clearly marked as a draft, viewable and downloadable, with the findings alongside.

Resumption, rework and report

Production runs server-side; the browser window may be closed. A job list shows all jobs with status and the phase reached. Persisted are the gap analysis, concept graph, detailed plans, chapter scripts, the output of individual text sections and every finished lesson — an interrupted run resumes where it stood.

If a chapter falls short, it can be regenerated with a hint. On completion a technical report is available: model usage per role with calls, tokens and truncated responses; scope by lessons, blocks and words; distribution of building blocks; density figures such as the share of non-textual blocks and words per interaction; and the production balance with lost lessons, degraded blocks and source coverage.