Overview
A generic AI writes generic posts. DemandBird's Voice Engine exists so the drafts you get back read like you wrote them: your sentence length, your habits, your opinions, the things you would never say.
It works in two halves. You feed it sources (an interview, past posts, a description of your writing). It turns those into a voice spec: a structured, readable, editable profile that every AI feature in DemandBird reads from. Nothing you write is rewritten to match it. The spec only shapes what the AI produces.
Find it at Settings → Voice. The top of the page shows how many sources you have filled in, and each source is a card you can start, continue, or re-run.
The interview
The interview is the fastest way to give DemandBird something real to work with, and it is the first step of onboarding for new accounts. Six questions, one per screen, at Settings → Voice → Interview.
- What do you do?Your role, business, services. Plain language, the way you'd describe it to a curious person at a bar. Feeds your positioning.
- Who are you actually writing for?Not a segment: the specific reader in your head, the one you'd send the post to first. Feeds your audience.
- What do you believe that not everyone does?Stances, not topics. “Pricing should be public” beats “pricing.” Feeds your themes and beliefs.
- What have you learned that surprised you?The thing you wish you'd known earlier, or that other people still don't get. Also feeds themes and beliefs, and it is one of the two inputs your content pillars are generated from.
- What's going on in your work right now?Current focus, recent wins, things in motion, things retired. Concrete names help. This is what keeps drafts from referring to a product you shut down last year.
- What in writing makes you cringe?Optional, and the highest-value question on the list. Specific beats abstract: “Hook, bullet list, ‘What do you think?’” beats “bad LinkedIn posts.” Becomes a hard never-do rule.
The first two questions are the floor: answer those and you have a usable voice. The rest sharpen it. Answers save as you go, so you can leave and come back, and the page picks up at the first question you haven't answered.
Talking instead of typing
Every question has a microphone. Record your answer out loud and DemandBird transcribes it, strips the “um”s and false starts, and drops the clean text into the box for you to edit. This is usually the better option: people write stiff and talk like themselves, and the interview is trying to capture the second one.
Transcription is biased toward your own vocabulary, so unusual product and brand names in later answers come out spelled the way you spelled them earlier.
Finishing and restarting
When you finish, DemandBird reads all your answers at once and builds the voice spec. You can restart the interview later from the Voice page; new answers replace the old ones and the spec is rebuilt.
Voice sources
The Voice page shows six source cards. Four are live today, two are marked coming soon. More sources means a sharper spec, but they are not all equal: the interview and your real past writing carry the most weight.
| Source | What it is |
|---|---|
| Interview | The six questions above. The one source you should always do. |
| AI description | If ChatGPT or Claude already knows how you write, ask it to describe your writing style and paste the answer in. DemandBird extracts the useful parts. |
| LinkedIn corpus | Import your past LinkedIn posts so the spec is learned from real writing rather than what you say about your writing. Request the export from LinkedIn and upload it; LinkedIn emails it about 24 hours later. The same file also unlocks the full LinkedIn analytics archive. |
| Published posts | Automatic. When you edit an AI draft and publish it, DemandBird saves the version you actually shipped as a voice sample. Your edits are the strongest signal there is, so this one improves on its own. |
| Pasted samples | Coming soon: paste a chunk of your own writing directly. |
| Handle lookup | Coming soon: pull your recent public posts from a social handle. |
The voice spec
Below the source cards, the Voice page shows the profile DemandBird built. It is deliberately readable: you can check it, disagree with it, and correct it. Sections:
- Identity: your positioning, your audience, and your current state (active products, secondary projects, retired things, communities, key collaborators).
- Voice signature: register, density, warmth, certainty, humor, perspective. Learned from your writing rather than stated.
- Signature phrases, opener patterns, pet metaphors: the specific moves that recur in your posts.
- Themes and beliefs: the stances you argue for.
- Never does: hard rules. Anything you named as cringe-worthy lands here.
- Mentioned entities: the people, products, and companies you name.
Each row shows where it came from, so you can tell a fact you typed in the interview from a pattern that was inferred out of your posts.
Editing and re-extracting
Identity, Never does, and Mentioned entities are directly editable: click a value, change it, save. An edited row is marked as yours and is protected from being overwritten the next time DemandBird re-reads your sources. Reset the row and it goes back to the extracted value.
Voice signature, signature phrases, opener patterns, and themes are read-only. They are observations about how you write, and the way to change them is to change what DemandBird is reading: add a source, or re-run extraction after you have published more posts.
Voice vs. brand
Two different things stack on every draft. Voice is yours and follows you everywhere. Brand belongs to the workspace: the positioning, the ideal customer, what is going on right now, and the content pillars.
That split is what makes agency work possible. One ghostwriter, one voice, five client workspaces: switch workspace and the drafts keep your craft but carry that client's brand. Brand fields are edited in Settings → Content topics, not here.
Voice and your AI limits
Building or rebuilding a voice profile does not spend your monthly AI allowance. The three metered buckets are drafting, rewrites, and post idea generation, and voice extraction is none of them: it is setup work, and charging you for it would punish the one thing that makes every later draft better. Run the interview, import your corpus, and re-extract as often as you like.
The caps that do apply live on the surfaces that generate copy. See Compose in depth for drafting and rewrites, and Content pillars for idea generation.
Team voice
Account admins get Settings → Team voice: every member of the account, how many voice sources each has filled in, and when their profile was last rebuilt. It answers the question “who on my team has actually set this up?” without asking them.
From the list, Manage voice opens that person's Voice page so an admin can run the interview with them, paste in an AI description, or fix a wrong fact in their spec. Non-admins only ever see their own.
If drafts don't sound like you
Work down this list in order. The first two fix most cases.
- Answer the cringe question. Negative constraints are stronger than positive ones. Name the specific move you hate.
- Import the LinkedIn corpus. Described voice is always weaker than demonstrated voice.
- Check Identity for stale facts. A draft that mentions a product you retired is an identity problem, not a voice problem. Edit the row.
- Keep editing before you publish. Every AI draft you rework and ship gets saved as a sample, so the profile tightens as you use it.
- Re-run extraction after adding a source, so the spec picks it up.