A subscription news outlet has 15,000 archived articles all written in a very distinctive, consistent editorial voice. They want a Foundry-hosted model to draft new articles that reliably match this voice without needing the voice guidelines repeated in every prompt.

The correct approach is fine-tuning a base model on the 15,000 archived articles so the voice is learned into the model's weights. With 15,000 consistent examples of a distinctive editorial voice available, fine-tuning lets the voice pattern be learned directly into the model's weights, so new drafts reliably reflect the style without repeating instructions in every prompt, which is exactly the scale of example set fine-tuning is suited for. The short system message describing the voice in three sentences is incorrect because it can nudge tone somewhat, but three sentences cannot capture the nuance of a voice built from 15,000 examples, and the outlet explicitly wants reliability without repeating guidance each time. The retrieval-augmented generation approach that pulls random archived articles in as citations is incorrect because retrieval-augmented generation is meant to ground factual answers in specific source documents for citation, not to teach a model a general writing style; pulling in random articles as citations would not make new, unrelated drafts sound consistently on-voice. The increased temperature setting to produce more varied phrasing is incorrect because raising temperature increases randomness and variety in output, which works against the goal of consistent, reliable voice matching rather than supporting it.