Sebastian Schiller-Stoff

Selected writing AI coding

EN — current language DE — Deutsch

Selected writing · AI coding

Autocomplete was never the point.

The 2023 Copilot workshop began with completions and ended with domain modelling. That order still explains why the tools work — and what they cannot decide for you.

Source record · 7 December 2023 · CC BY 4.0

Why it exists

In December 2023, Copilot had just become a product with a price.

The workshop at the Centre for Information Modelling in Graz was split in two: a theory part by Julia Martin on what Copilot is, what it costs and what is behind it; a hands-on part by me in Visual Studio Code. The room held people who had never installed an extension and people who wrote code daily.

The question was not whether the tool was impressive. It was whether a research group could adopt it without losing the ability to explain its own software.

The workshop’s answer was structural: these tools do not remove design. They move it.

The argument

Four surfaces, one shift.

Copilot in 2023 was four things in an editor: autocomplete, a completions panel, slash commands that operate on the open file or the selection, and a chat. The workshop demonstrated all four, then spent the rest of its time on what they change.

The first change is speed on introductory problems: boilerplate, a class, an explanation of unfamiliar code. The second is risk: the training material is opaque, and the tool is confident about code it has not verified.

The third change is the interesting one. If the assistant writes the technical layer, the human work moves to the domain layer. The workshop called it a comeback of object-oriented modelling: name the real problem, not the technology. A class called Student is a claim about the domain; a class called ServiceLayer is a claim about the framework.

Two theses followed. First, code that follows the domain produces better results with these tools. Second, assistants shift software design toward domain modelling and relieve repetitive work — provided the domain is modelled adequately. The proviso is the whole sentence.

  1. Autocomplete

    A suggestion at the cursor.

  2. Completions

    A panel of alternatives to review.

  3. Commands

    An action on the open file or selection.

  4. Chat

    A conversation with the repository in scope.

What holds in 2026

The surfaces survived; the centre moved.

Autocomplete is still there, but the interesting unit is now the change set: a diff across files that a person reviews. The 2023 deck treated completion as the centre because that was the product. By 2026 the product is the edit.

The pricing slide is obsolete — the numbers changed, the education licence stayed. The lineage held: search, then completion, then chat, then the command line.

The domain-modelling thesis held better than the productivity claim. A tool that writes the technical layer makes a bad domain model more expensive, not less.

What I would change: start with the diff, not the completion. And spend more time on the review gate, which in 2023 was a slide and in 2026 is the job.

The record

The deposit is the workshop deck: one PDF, with the tool comparison and the modelling argument.

Title
Copilot technologies in the digital humanities
Type
Presentation
Deposited
DOI
10.5281/zenodo.10288101
Co-author
Julia Martin
Licence
CC BY 4.0
Reach
862 downloads, 770 of them unique (Zenodo, 15 November 2026)
Contents
workshop slides (PDF)

Colophon

Method
A reading, 15 November 2026. It describes what the workshop demonstrated and what I would teach differently now; it is not a reprint.
Sources
The 2023 workshop deposit and its slides, read against the 2025 FAIR-code paper and the 2025 AI-coding workshop.
Authorship
Sebastian David Schiller-Stoff. The workshop was co-authored with Julia Martin. The reading is mine.
JavaScript
None. The page is text, CSS and a small inline sprite.
Licence
Text and layout: CC BY-NC 4.0. The cited record keeps its own CC BY 4.0.

All writing Back to the CV