Practice note · AI coding assistants
The assistant writes code. You sign it.
Since 2023 I have taught research teams to use coding assistants without handing over judgement: what the tools do, where they fail, and what a reviewed AI change has to pass. This page is the workshop record, the case behind it, and the limits I teach with it.
What changed
In 2023 the question was whether autocomplete could finish a line; by 2024 whether a chat window could refactor a module; by 2025 the tools could read a repository and run a command.
The interface moved; the risk stayed the same: code that looks right in the editor and is wrong in the system. Teach the tools by their failure modes, not their feature lists.
What the tool is good at
- Boilerplate
- Test scaffolding
- Unfamiliar APIs
- Explaining legacy code
What it cannot see
- The research question
- The data model
- The licence
- The consequence of a publish
The curriculum
A workshop series, not a one-off demo. December 2023 to September 2025: autocomplete → chat → refactoring → review. Every step is open on Zenodo.
- Kann LLM generierter Code FAIR sein? Herausforderungen für das Forschungsdatenmanagement
- Der Einfluss von AI-Pair-Programmers auf die Digital Humanities: Potentiale und Limitationen
- KI gestützte Programmierung in den Digitalen Geisteswissenschaften
- AI coding assistants (advanced)
- AI coding assistants
- IDE-Integration of GitHub Copilot: Jetbrains IDEs
- Project Overhaul und Refactoring der digitalen Edition der „Urfehdebücher der Stadt Basel“ mithilfe von GPT-4 und LLM
- Legacy code and copilot technologies: AI-supported software maintenance
- Copilot technologies in the digital humanities
The series was taught with Münzer, Sagadin, Dittmann, Martin, Pollin, Scholger, Steiner, Lang and Galka.
The case
The Urfehdebücher der Stadt Basel edition needed an overhaul: data model, transformation, visualisation and the web application.
With the UFBAS team we used GPT-4 across those steps — TEI XML, categories and SKOS, an ontology, RDF, frontend work and data visualisation — and documented where it helped and where it did not.
The lesson was not that the model wrote the edition. It compressed the distance between a decision and its implementation, which made review the bottleneck and the most valuable part of the process.
What the model did
- Draft transformations
- Propose mappings
- Generate frontend scaffolding
- Explain unfamiliar code
What the team decided
- The data model
- Ontology commitments
- The licence
- What counted as a correct transformation
- What shipped
The limits
The 2025 FAIR-code abstract names the failure modes: generated code with poor maintainability, missing metadata and licence uncertainty is hard to find and hard to reuse lawfully.
Insecure libraries and inconsistent standards compound it. In research that is a data-quality problem, not a style problem. Vibe coding is a starting point, not a method.
-
The diff is small enough to read
Review is a reading task, not a glance.
-
The tests ran
The check is a command, not a promise.
-
The dependency and its licence are known
Every new library is a legal and maintenance commitment.
-
The prompt and sources are recorded
Provenance is part of the artefact.
-
A person signs the commit
The signature is the method.
The signature
The assistant can write the code. It cannot sign.
The brief, the test, the licence and the signature stay with the person who ships. That is the whole method, and it is why the record above is a teaching record rather than a tool review.