Mass content production with agents

Generating one article with a model is easy. Producing two hundred pieces a month with consistent quality, in three languages, with voice and video, is a systems problem.

Here is how I structure a content factory.

Stages

brief → research → outline → draft → review → voice → render → localize → publish

Each stage is an agent with one job and one output format. The output of one stage is the input of the next. Nothing goes forward without passing a check.

The checks matter more than the prompts

Prompts get all the attention. Checks are where quality lives.

• Factual check. Claims are traced to sources gathered in the research stage. Unsupported claims are removed or flagged.
• Voice check. A style guide, expressed as examples, is compared against the draft. Drift gets rewritten.
• Duplicate check. New pieces are compared against everything already published.
• Format check. Length, structure, metadata, links. Boring, and the most common failure.

A piece that fails goes back one stage with a note. A piece that fails twice goes to a human.

Voice and video

Text is the cheap part. Voice with ElevenLabs and video rendering turn one script into several formats. The trick is to generate the script with the final formats in mind: short sentences, no visual references that only work in text, pronunciation hints for names.

Localization

Translation is its own stage, with its own review. A literal translation of a good English piece is a bad Portuguese piece. The localization agent gets the brief and the sources, not just the text, so it can rewrite rather than translate.

Publishing

The last stage talks to the CMS, the video platform and the social accounts. It schedules, it does not post immediately. A human can still pull something before it goes out. In practice they rarely do.

The result

A team of one runs a content operation that used to need six. Not because the model is a great writer, but because the pipeline never forgets a step.