Teaching the AI your voice

Tell it directly

Say things like "I come across too formal", "stop using the word 'journey'", or "we launched a new service called X" right in the chat. The AI folds these into your brand voice document immediately and confirms what changed.

Retrain vs. Redo interview

Two similarly-named buttons in Settings → Brand Voice do very different things:

• Retrain rebuilds your brand voice document from everything you've already provided — interview answers, writing samples, website analysis, brand context. Nothing you've entered is touched; it just regenerates the document (and bumps its version). Use it after updating your inputs, e.g. a new writing sample or website.
• Redo interview asks the interview questions again. Nothing is deleted when you start it: your current voice document and answers stay in use until you finish the new interview, so stopping halfway leaves everything as it was. Use it when your business or positioning has genuinely changed and the old answers no longer describe you.

Rule of thumb: new inputs → Retrain. New you → Redo interview.

Neither of them deletes anything. If what you want is a clean slate — the drafts, the chats, the projects and the learning all cleared — that's Start this brand over in Settings → Brand. See Starting a brand over.

It learns from your edits

Once a week, SocialStache compares what the AI drafted against what you actually published — your edits are the signal. Recurring patterns get folded back into your brand voice so drafts drift steadily closer to how you'd write them yourself.

It won't make up facts about you

The AI writes in your voice, and a first-person post wants specifics — a date, how long you've been doing this, who does what. It will only use ones you've actually given it, in the interview, your brand context, or in chat.

Where a post needs a detail it doesn't have, it writes around it or tells you what's missing. It won't fill the gap with something plausible, because a plausible invention is the kind nobody checks — and it goes out under your name.

If something slips through anyway, we catch what we can: a post claiming a date, a team history, or a customer count we were never told gets flagged above the caption with "Check these before posting". That check can't know whether something is true, only whether you told us — so it's a prompt to look, not a verdict.