This repository is the home of a teaching line on working with large language models in research and in professional knowledge work. It covers what a language model is and where its capability is uneven, how a prompt addresses it, how project knowledge and working context are built and maintained, how multi-step agentic work is organized and controlled, and how the results are checked, governed and accepted. The audience is general, everyone who does computer-based, data-driven knowledge work.
The material is maintained through three holdings, each with its own function. The Full Slide Deck is the complete overview of the teaching material in slide form. It serves as a reference work and as the source from which workshop-specific decks are assembled, and it is not presented as one teaching session. The Full Lecture Notes are the written-out counterpart of the slides and carry the explanations and context needed for reading and later reference. The Video Playlist combines DHCraft videos with selected contributions by others; its deliberate sequence runs from the foundations to more complex topics and forms a structured learning path that is curated, maintained and extended as suitable material appears. All three holdings share the same five modules, and the playlist assigns every entry one primary module.
The public learning architecture keeps two general entry points: the landing page and Preparation. Preparation combines a verifiable technical setup with four compact introductions. The five numbered modules remain the structure of the Full Slide Deck and the classification used for workshops and videos; they no longer create five additional overview pages.
| Module | Title |
|---|---|
| 1 | Understanding Large Language Models |
| 2 | Prompt Engineering |
| 3 | Knowledge and Context Engineering |
| 4 | Agentic Engineering |
| 5 | Critical Perspectives and Governance |
Every workshop taught from this material is a documented specialization. Its profile selects modules, sets depth, case studies and language, and adds what belongs to that one instance, the schedule, the hands-on instructions and the venue specifics. An improvement to a module therefore reaches every workshop that uses the module, and the core of the material stays domain-neutral because the case studies are exchangeable modules of their own.
Only upcoming instances live in this repository. Once an instance has been taught, its state is tagged and its folder leaves the working tree, so the platform shows what people can still attend. The material stays retrievable in git history under its tag.
| Path | Content |
|---|---|
knowledge/ |
Project knowledge base (Promptotyping documents); start at knowledge/INDEX.md |
script/full-lecture-notes-en.md, script/full-lecture-notes-de.md |
The Full Lecture Notes |
script/modules/ |
The five-module cut of the lecture notes, with its coverage report in script/COVERAGE.md |
slides/full-slide-deck.md |
The Full Slide Deck as text with speaker notes; the visual deck lives in Google Slides and is linked |
docs/data/videos.json |
The Video Playlist inventory, including order, primary module, any related modules and duration |
workshops/<id>/ |
One folder per upcoming instance, holding profile.md, slide-deck.md, lecture-notes.md and data/ |
docs/ |
The platform (static site, GitHub Pages), including the register docs/data/workshops.json and the covers in docs/assets/covers/ |
docs/preparation.html, docs/preparation.de.html |
Technical setup and four compact subject introductions |
docs/tutorials/ |
Bilingual AI Harness introductions for Claude Code, Codex and Pi |
communication/ |
The Research Mission Control transport for current briefs, results and recipient state |
Instance ids have the form YYYY-MM-DD-slug and are the same string in the register, the workshop folder, the cover file, the subpage and the git tag.
The platform is a no-build static site in docs/; its content and navigation require no JavaScript. The active tree contains 26 HTML files in thirteen English/German pairs. The landing page presents the five Full Slide Deck modules as a non-interactive contents index. Compact workshop markers lead to the stable landing-page module anchors. Preparation carries four compact subject introductions whose public cards contain only accepted labels and routes until their slide-derived teaching content has been reviewed.
After Workshops, the landing pages place Preparation beside four compact learning routes. AI Coding Literacy is the available external German curriculum. Claude Code, Codex and Pi are local bilingual routes from the Agentic Engineering introduction. Research Data and Knowledge Work belongs to the maintained slides and is not a separate route. The three Harness introductions use a comparable eight-step sequence from purpose and prerequisites to a first reversible check.
Release v0.3 is preserved at commit aa4193378413a9723a2144d155b186b933229435 and tag v0.3. The active working tree supersedes its module-hub topology through the Research Director decision of 26 August 2026. The simplified state has passed local formal and browser verification; Research Director acceptance and release remain open.
https://dhcraft.org/knowledge-context-agentic-engineering/
Promptotyping is the artefact-development method. Maintained documents in knowledge/ guide iterative production, inspection and correction of teaching material, interface code and workshop packages.
Research Mission Control is the orchestration contract for accountable research and implementation lanes. The Research Director retains substantive authority. The Research Orchestrator frames research and issues bounded implementation briefs. The Operational Orchestrator coordinates authorized implementation, verification, integration and release. The Context Stream in communication/ transports current briefs, results and recipient state; durable project knowledge remains in knowledge/.
The concepts, arguments and editorial decisions in this repository originate with Christopher Pollin. Almost all prose, interface copy, diagrams and implementation artefacts were drafted with AI harnesses and revised through repeated Promptotyping and Research Mission Control cycles. Human authorship here means responsibility for the ideas, source selection, decisions, evaluation and publication; it does not imply manual production of every sentence or file. Third-party sources and teaching materials retain their own stated authorship and rights.
The condensed decision record is maintained in knowledge/journal.md. The complete sequence of revisions, including superseded formulations and technical iterations, remains inspectable in the Git history.
Repository files are English. The teaching material is bilingual, and language is handled per module and per instance, so each holding carries the language it was authored or taught in.
Text content is licensed CC BY 4.0. Slide decks are checked individually for third-party image rights before the license statement is extended to them. The executable CLARIAH-AT hands-on package lives in zbz-ocr-tei.
Christopher Pollin, Digital Humanities Craft