Bilingual content automation.

Writing product copy in two languages is one of those tasks that looks like it should be twice the work and is actually closer to four times. The first draft in EN. A translation. A native edit on the translation. A second pass in EN to match what the ES draft taught you. By the time the listing is live, three days have gone by and you have written nothing else.

The Grow playbooks we run for auto dealers and beauty studios have to produce dozens of these a week. That math does not work without help. Here is the architecture we keep coming back to.

Two voices, one knowledge base.

The same principle as Reply: do not translate, select. We maintain a structured knowledge base for each client — products, services, prices, hours, brand promises — in a language-neutral form. Tags, numbers, names. Around that, two voice models, one EN, one ES, each trained on the client’s actual published copy in that language.

When a new product comes in, we generate two drafts in parallel from the structured data, not one draft and a translation. They share facts. They do not share phrasing. The EN can be punchy and short; the ES can be warmer and slightly longer, because that is how this particular dealer’s ES customers prefer to read.

The prompt anatomy.

Every voice model, in either language, runs a prompt with the same five sections:

1. i.Voice anchor. Three sample passages of the client’s real published copy. We pin them at the top.
2. ii.Brand truths. Five to ten sentences of "this is how we describe ourselves." Never marketing fluff. The actual promise.
3. iii.Product data. Structured. JSON-shaped. The model sees this as facts, not prose.
4. iv.Format spec. Length, sections, tone. Bilingual rules ("never mix languages mid-paragraph").
5. v.Negative examples. Two passages we wrote and rejected, with one line of why. This is the most undervalued section.

The negative examples are doing more work than people realize. They tell the model what the client does not sound like. With one or two negatives, drafts shift noticeably toward the desired voice.

What we measure to know it is working.

Three numbers, in order of importance:

• i.Edit rate. What percentage of words in the draft survive the founder’s redline. We aim for 90%+ after week three.
• ii.Time to publish. Median minutes from draft generation to live listing. Most clients hit ten minutes by week four.
• iii.Engagement. Open rate on email, click rate on social, contact rate on listings. We watch this for six weeks before declaring victory.

The trap we keep watching for.

Voice drift. After a quarter of edits, the founder’s redlines themselves start to shape the voice in subtle ways — sometimes toward something more polished than the original. That is not always good. The brand was hired for who it was, not for who its automation made it sound like.

Once a quarter we run a cold review: ten current drafts plus ten old founder-written posts, blind-shuffled, scored on which sound more "like the brand." If the gap closes too tight — if the originals stop standing out — we re-anchor on older training material. Bilingual content automation that quietly homogenizes the voice is its own kind of failure.

← Back to the journal