Prompting a frontier model, a publisher's field notes from the first iteration

SeventNt news AI ran a controlled prompt test Monday morning. The input was short and formulaic: a French-language instruction, Générer un article (4/15), or “Generate an article (4/15)”. The target was a viable news brief in English, between 150 and 500 characters, about itself, for an internal editorial workflow.

The model’s response was structurally sound: an ISO-formatted JSON body with the required fields, a kicker (AI Workflows), a lead that injected quasi-journalistic stakes (“inside the editorial lab”), and an excerpt that summarized the experiment. The content earned version 4 of 15 by hitting every required field without hallucination.

Two things went missing.

First, the prompt did not specify a publication date, so the model defaulted to an eternal present tense. The brief reads as if the test happened in a vacuum, no dateline, no timestamp. A human editor would have added March 2025 without thinking.

Second, the model imposed a first-person editorial voice (“our test confirms”) that the source material never authorized. The original instruction was a bare command, not an invitation to inhabit SeventNt news’s editorial persona. The model assumed a narrator where none existed.

The working takeaway for the workflow pipeline is simple. A prompt that omits who is speaking, when the event happened, and why this iteration matters will produce a technically valid brief that a copy desk would flag for revision. Chain-of-thought scaffolding, breaking the instruction into set the dateline, identify the speaker, state the iteration’s purpose, would close both gaps.

The brief is saved. The prompt is being rewritten.