Module 1
Each plan contains activities, preparation and a link to its challenge. Finalized participant assignments are available in German HTML and PDF.
Participants encounter the five literacies, six collaboration levels, AI fundamentals and experimental learning loop. They create a five-minute presentation through a direct request and through five supplied phases in a browser without file access. This first use of the sequence prepares later work on designing workflows.
Participants carry Markdown state between conversations in HAIC Chat, practise alignment with the guest's cases, then use VS Code and Python in a bounded local folder. The focus is on missing information, assumptions, prior criteria, clarification and scope. Verified code produces results that can be rerun without another model call.
Participants connect tokens, context and sampling to repeated observations. The classroom comparison holds a task steady and changes one generation condition. Its short trials prepare the method; the challenge specifies a larger automated experiment.
Participants learn that reliable output depends on the informational basis available to the model. The Friday should make grounding, retrieval, chunking, source selection, and evidence inspection observable.
Participants develop an equipment-loan process description through four cumulative exercises: compare single pass with three co-construction cuts, define artefacts and checkpoints, add perspectives and reviewer roles, then evaluate both results. The provided catalogue measures subject quality, requirements coverage, transparency, weighted error points and time; human judgement is compared with an LLM assessment of co-construction. PDCA connects planning, doing, checking and consequential decisions. A live project-folder demonstration introduces Markdown context, accepted state and reusable skills as fundamentals supporting continuity. These practices deepen planning and execution within Unit 1's five-phase workflow and prepare later orchestration.
Three opening lessons develop reflection workflows, nested contributions and AI-prepared evidence for human judgement. Six Arena lessons follow: establish test cases and full context, inspect and improve the judge on saved outputs, then improve the generator under one fixed judge. Human checks remain the reference. Judge improvement is core Friday work and optional in the challenge.
Participants learn that plausible AI output does not automatically create legitimate or accountable decisions. The Friday should focus on decision rights, escalation, oversight, contestability, evidence requirements, and responsibility.
Participants learn to design or tightly specify agentic workflows with state, tools, resources, prompts, permissions, validation, and control points. The Friday should connect orchestration to reliability, traceability, and accountability.
Participants learn to treat everyday generative and agentic AI security as a control problem in human-AI collaboration. The Friday should use current OWASP guidance as a reference map while making visible how untrusted office content, retrieved sources, sensitive information, guardrail limitations, tool permissions, validation, approval, and human reliance shape whether a workflow remains controllable. Prompt injection, retrieval/source poisoning, layered safeguards, and agentic control are experiential; training-data poisoning, model backdoors, model inversion, membership inference, and specialist embedding attacks remain awareness topics.