Talk, Check, Decide: A Skill-Based Workflow for Writing and Verifying Geological Excursion Reports with an AI Assistant

Abstract. Large language models write fluent text but invent references and blur the line between what was seen and what was read. Geological excursion reports are sensitive to both failures, because every statement must be bound to a place and a source. This paper describes a workflow in which an AI assistant (Claude) works under a family of 20 written procedures, called skills, and two templates for its projects, that separate research, writing, checking and decision, and in which the author, by his own account, works almost entirely by voice. The workflow is shown on two reports: a post-trip report on 20 localities in Germany and a pre-trip field guide on 21 localities in southern Norway and Denmark. Six review axes are kept apart: fact support, internal consistency, reference accuracy, locality evidence, genre and usefulness. In the Norway report, seven consistency runs produced 18, 4, 6, 4, 0, 2 and 71 findings, the last after the author had changed the target format from a planning report to a dossier; in the Germany report, a full reference check found 13 errors in the 211 entries it held on 18 September 2026. In three documented cases, an error was caught by one axis only. The results suggest that the value of AI assistance lies above all in systematic, counted checking, with a human in the loop: the assistant proposes the complete wording, the author confirms it, and only then is it inserted unchanged, logged and verified.