Skip to content

Präsentationsvorbereitung

The application supports the structuring of talks and presentations from existing documents. From uploaded material — lecture notes, papers, reports, or existing slide decks — a structured Markdown artefact is generated, which can subsequently be exported as Markdown, PowerPoint, or Word. Generation does not rely on a single LLM call but proceeds through an agentic pipeline of two specialised agents that handle source fidelity, length adjustment, and translation in a controlled manner. The LLM connection is designed as an OpenAI-compatible interface and can be operated locally or in an organisation's own data centre.

At a glance

  • Derive a structured slide outline from existing documents (PDF, Word, PowerPoint, Markdown, text) without producing it manually
  • Adjust the number of slides to the planned talk duration, instead of having to shorten or extend it afterwards
  • Translate an existing presentation into another language, including speaker notes
  • Combine content from multiple source documents into a single coherent slide outline
  • Refine the generated structure through dialogue, without searching the source text again each time
  • Export the result in the format required (Markdown for further processing, PowerPoint for the talk, Word for accompanying documentation)

Highlights

In contrast to a direct LLM prompt or a simple conversion script, the application separates dialogue from structural work, checks every bullet point against the source documents, and adjusts the slide outline systematically to constraints such as duration and language. The resulting differentiating characteristics are as follows:

  • Two-agent architecture: A chat agent conducts the conversation and gathers requirements; an artefact agent processes each dialogue turn separately and updates the slide outline within a structured Markdown document. Both agents have separate system prompts and distinct responsibilities.
  • Source fidelity with internal citations: Content slides are derived exclusively from the uploaded documents. During generation each bullet point carries an internal source reference; missing information is marked as an explicit placeholder. Inventions on content slides are excluded by the prompt rule set.
  • Time-based slide adjustment: A target slide count is calculated from the stated talk duration (factor 0.8 slides per minute). A second processing phase reduces or expands the outline specifically to this target, instead of producing an arbitrary slide count.
  • Multilingual processing: When the target language differs, headings, bullet points, and speaker notes are translated in full; technical terms can be glossed in parentheses where appropriate. The language switch is also possible later within the ongoing dialogue.
  • Connection to five input formats: PDF, Word (DOCX), PowerPoint (PPTX), Markdown, and text are read through a unified parsing component and mapped to a common Markdown representation.
  • Three output formats: The output can be downloaded as Markdown, as a PowerPoint file with speaker notes, or as a Word document with slide-level sections.
  • Version history with undo: The most recent five artefact versions are retained; each change is given a short diff description and can be reverted.
  • Token budget with visibility: The combined token usage of documents, dialogue, and artefact is shown continuously in order to stay within the context limit of the LLM in use.
  • Local LLM connection: The application expects an OpenAI-compatible interface that can be operated locally or within a private network. Content therefore need not leave the operator's own infrastructure.
  • Session-based operation: Data is held only in the memory of the running session; nothing remains in the application after the session ends.