Interviewing yourself

AI is relatively good at producing long documents, which is powerful; but left to itself it tends to write in a sloppy, generic way. The usual fix is a writing guideline: a document that describes how the writing should actually be done. I have two: one covers my Research Musings substack, which, aside from tone, asks that the writing stays focused on the methodology rather than the results, and never overstate what those results show. The analyst version goes further: it also covers colours, charts (most of Tufte's visualisation rules, with my extension: the bumped stacked bar), and tone. Rather than write either from scratch, I asked a model to interview me.

Below is the interview prompt I used for Research Musings. The interview had three parts: 1. focused questions about how I know a piece is finished, which posts had failed, and what I expect the reader to do with the argument; 2. a sentence-level taste test, where five short passages interpreting the same invented bibliometric finding in different registers had to be ranked and explained; 3. and a final round where the model named contradictions between my stated preferences and the archive, and asked me to resolve them.

The interview helped me to express how I write. Some answers confirmed things I already knew, such as the need for a post to leave the reader with both a methodological and a substantive takeaway (the first part of each Research musings). But it would have taken me longer to write that my posts usually use the empirical example as a way of showing a method (AI tends to favour the results), that I prefer sentences which hold an entire argument instead of slices in multiple sentences, and that humour is allowed only when it is directed at myself or at something faintly absurd in the material. The discussion unearthed this, and the use of examples helped making sure my comments were contextualised.