A Lyon News Desk Ships French Summaries Before Noon

The Lyon desk has three summary jobs before lunch

Picture a small news desk in Lyon monitoring city council meetings, transport updates, court reports, and local business announcements. By noon, editors need short French-language summaries for a website, a mobile alert, and an internal morning brief.

The difficult part is not asking an AI model to summarize one article. The difficult part is keeping the work predictable when every item arrives in a different format and carries a different level of risk. A transcript of a routine meeting can be handled one way. A legally sensitive report needs a slower review. A mobile alert needs a strict word limit.

The useful starting idea is simple: assign one model to each recurring job. Do not ask one general-purpose workflow to make every decision.

Fast first-pass summaries use Claude Haiku 4.5

The desk sends clean source text to Claude Haiku 4.5 for the first pass. Its job is narrow: identify the main event, extract named places and organizations, and produce a short French summary in a fixed structure.

A practical instruction might require a headline, a two-sentence summary, and a list of facts that need checking. The prompt should also say what the model must not do: invent quotes, merge separate events, or turn uncertainty into a definite claim.

This model is a good fit for high-volume material such as agenda items, routine announcements, and transcripts that have already been cleaned. The editor receives a consistent draft quickly, while the system records the original source beside it.

Nuanced stories use Claude Sonnet 5.5

A second job needs a different model. Claude Sonnet 5.5 can handle a more careful editorial pass for stories where context and wording matter. The desk can give it the source, the first summary, a glossary of local names, and the publication format.

Its task is not to silently rewrite everything. It should compare the draft with the source, flag unsupported statements, preserve attribution, and suggest a clearer French version when the original wording is dense. This is especially useful for stories involving policy changes, conflicting statements, or technical subjects that readers may misunderstand.

Keeping this as a separate job gives editors a visible decision point. A fast draft can move through the queue, while higher-risk items receive a more deliberate review. The model is assigned because of the job requirements, not because it happens to be the default in an application.

A routing table keeps the newsroom predictable

• Routine announcement: extract facts, write a short French summary, and mark it ready for editorial review.
• Long transcript: remove repetition, identify the central decision, and preserve speaker attribution.
• Sensitive report: compare every important claim with the source and list unresolved ambiguities.
• Mobile alert: produce one concise sentence without adding information that is absent from the source.
• Publication check: verify names, dates, places, numbers, and the requested word limit.

Each row should define the input format, the assigned model, the output schema, and the human approval step. That turns model selection into an operational rule that can be tested and changed.

French quality depends on editorial controls

Language quality is more than correct grammar. A Lyon audience may care about whether a commune, district, transport line, or public body is named precisely. Prompts should include approved spellings and tell the model to keep French punctuation and register consistent with the publication.

Use structured outputs where possible, with separate fields for the summary, source facts, uncertainty notes, and review status. Store the source text and generated result together. Editors can then see whether a mistake came from extraction, interpretation, or a later rewrite.

RelayVano fits a pay-as-you-go newsroom workflow

For a small team, RelayVano is a practical way to test this pattern without committing to a single provider contract. It gives developers one relay for multiple AI models, works with OpenAI-compatible SDKs and tools such as Claude Code, Codex, and Gemini CLI, and uses the API base https://api.relayvano.com. You can route each job through the model selected for that job and track usage as the desk grows.

Start with two queues, measure approval time and correction rate for a week, then adjust the routing table. Sign up for RelayVano when you are ready to connect the workflow and run the first batch of French summaries.