This post was written by the tool it describes. Here is the brief.

Six fields

This post started as a brief with six fields: topic, angle, audience, language, tone and length. The topic was this post. The angle was the trace: show what happened between the brief and the approved version, step by step. Audience was general; language English; tone neutral; length standard, 1,000 words.

Nothing else was typed. No outline, no keywords, no sample paragraph. The brief is the article in the sense that everything the pipeline did follows from those six fields and the source set they produced.

What the pipeline did

Search found seven outlets writing about how newsrooms use language models for drafting. Grouping merged three of them, all carrying one agency report, into one origin, which left four independent origins to read. Each was read once. Nine claims were extracted from the draft and checked against those four; all nine were corroborated by at least two of them, so this post has no underline.

The write step produced 1,020 words in 38 seconds. A second model, not the one that wrote, read the draft back for neutrality and for similarity to the sources. The closest passage was 9 percent similar to its source, below the threshold, so no rewrite was asked for.

What the editor changed

Two sentences were rewritten by hand: the opening, which was accurate and dull, and a transition that repeated the previous paragraph. The editor approved version 2. The label under the source list is derived from that approval record; it cannot be set by hand, which is why it reads the same on every post here.

Frequently asked questions