8.6 Copy, Comments, and Community
Captions and comments are not decoration. They are where the persona's voice becomes verifiable: an audience decides whether an account is a person or a generator largely from how it writes in twenty words.
They are also the cheapest research instrument available, and the place where you will first be accused of being what you are.
Captions in the persona's voice
At sentence level, the The six-part specification that makes a persona reproducible as a character across months. (2.3) has to become rules you can apply while writing under deadline. Four that carry most of the effect:
Sentence length. If the persona speaks in short sentences, the captions are short. A persona whose voice is clipped in conversation and writes in long subordinate clauses in captions reads as two entities.
Vocabulary boundaries. Words the persona uses and words she never uses — from the bible's voice section. The banned list from STYLE.md §5 applies here too, and for the same reason: certain words are generic enough that their presence signals a template.
Punctuation habits. Whether she uses em dashes, ellipses, exclamation marks, lowercase. These are more identifying than vocabulary and easier to hold consistently.
What she does not do. No motivational filler, no engagement bait ("comment YES if..."), no brand-speak. Each of these makes an account read as marketing, and the audience's tolerance for it from a persona is lower than for a human creator, because the suspicion is already present.
One practical method: write the caption, then read it as if it were a comment from someone else. Would you believe a specific person wrote it? This is the same test as 6.6's read-it-aloud check, applied to text.
Comments as free research
The comment section is the only place an audience tells you what it wants, in its own words. Four things to harvest, and where each is used.
| What you find | What it is | Where it goes |
|---|---|---|
| Repeated questions | Topics the audience wants covered | The next content plan (2.6) |
| Complaints | Unmet expectations, format problems | The format test in 8.2 |
| Requests | Specific content asks, sometimes purchasable | Product ideas (9.6) |
| Language | The exact words the audience uses | Captions and search terms (8.4) |
The method. Once a week, in the review slot (7.6), read the comments on the top five posts and write down the questions that repeat. Ten minutes, and it produces better topic input than any brainstorming session, because it is the audience's own list rather than your guess.
The fourth row is underused. The audience's actual vocabulary is what makes captions and search terms work. If they say "small apartment" and you write "compact urban dwelling," you have written for nobody.
Moderating the accusation
The conversation you will have, usually in the first weeks, and it should be scripted rather than improvised.
Three prepared responses, for three situations.
A direct, good-faith question ("is this AI?"). Answer plainly and without defensiveness. One or two sentences, in the persona's voice, confirming the nature of the account and what it is for. The disclosure templates in 11.3 cover the wording; the tone is the persona's, not a legal disclaimer's.
A hostile accusation ("this is fake, this isn't a real person"). Answer once, calmly, without argument, and do not continue. The audience reading the exchange is who you are answering for, and a persona that argues with a stranger looks worse than one that acknowledges and moves on. Do not delete a fair question; delete harassment and slurs per your moderation policy.
A bad-faith pile-on. Not a conversation; a moderation task. Delete what violates the platform's rules, block where appropriate, and do not attempt to win. A pinned disclosure (8.2) reduces how often this happens, because the question is answered before it is asked.
What not to do, in any case: deny being synthetic, claim to be a real person, attack the accuser, or delete the question while leaving the accusation unanswered. The first two are the conduct that turns a labeling issue into a deception issue (11.4); the last is the one that makes it worse.
Community without pretending
The premise worth stating: an audience can know the persona is synthetic and still form an attachment to it. The trust numbers from 1.4 are lower for virtual creators, and the engagement numbers are higher, and both are consistent with the same thing — the audience relates to a character, with the reservation that it is a character.
What this means for community building:
Do not build the community around the claim of being human. It is a foundation that fails, and the failure is public.
Build it around the persona's domain. The recipes, the routines, the niche, the aesthetic. That is what an audience can engage with, and it survives the disclosure question entirely because it never depended on the answer.
Let the persona have a consistent voice and opinions. Character is what attachment attaches to. A disclosed persona with a strong character outperforms an undisclosed one with none, and it is not close.
Be consistent about the disclosure. Every post, every bio, every pinned introduction (11.3). The audience notices inconsistency more than it notices disclosure.
The transparency phase
When to disclose, and how to think about the question.
The legal floor is not the same as the strategic question. Disclosure obligations come from platform rules and, in the EU, from the AI Act's transparency requirements (11.3 covers the specifics). Those are non-negotiable and they are not a phase.
What is a genuine strategic question is whether the account's positioning foregrounds the synthetic nature. Three viable positions:
Minimal. Disclosed as required, in the places the rules require, with no emphasis. The account is about the niche; the nature is stated where it must be.
Open. The synthetic nature is part of the account's identity and is mentioned where relevant. Some audiences find this more interesting than the content itself, which cuts both ways.
The subject. The account is explicitly about synthetic personas — building them, living as one, the process. A different product with a different audience, and it is a legitimate niche.
How to decide. By what the audience came for. If they came for the niche, minimal disclosure keeps the account's center of gravity where the value is. If they came for the novelty, foregrounding it extends the interest.
And one practical note about transitions. If you move from minimal to open, do it deliberately and once, with a post that makes it part of the story (2.4's planned events). A gradual drift of disclosure emphasis reads as an account that changed its mind under pressure.
Two things that are not optional
Disclosure stays on every post. The positioning question above is about emphasis, not about whether. Every platform's requirement and, in the EU, the AI Act's transparency duties (11.3) apply regardless of which of the three positions you choose. [AS OF 2026-10] The rules tightened during the research period for this book and are still moving; verify before relying on a template.
The persona's voice survives handoff. If anyone else ever writes a caption or a comment reply (12.6), the sentence-level rules from the first section of this chapter are the handoff document. Voice is the cheapest thing to lose and the most visible when lost.
Where this goes wrong
Captions that do not match the voice in the bible. Two entities, one account.
Engagement bait from a persona. Lower tolerance than for a human creator.
Ignoring the comment section. The only place the audience says what it wants.
Writing in your own vocabulary. The audience's words are the searchable ones.
Denying being synthetic. Turns a labeling question into a deception problem.
Arguing with an accuser. The audience is watching, and the argument is the content.
Deleting a fair question. Leaves the accusation unanswered and the deletion visible.
Building the community on the claim of being human. A foundation that fails publicly.
Drifting disclosure emphasis without a decision. Reads as pressure, not position.
The gate
- Can you name four sentence-level rules from your bible that a caption must follow?
- Do you harvest repeated questions from comments weekly?
- Do you use the audience's own vocabulary in captions and search terms?
- Do you have three scripted responses to the AI accusation, including the good-faith question?
- Is your pinned introduction enough that the question rarely needs to be asked?
- Have you made a deliberate positioning decision on disclosure emphasis, rather than drifting?
What you leave with
Four sentence-level voice rules with the read-it-as-a-stranger test, a four-column comment harvesting table routed into the content plan, three scripted moderation responses with what never to do, the premise that a disclosed persona can still earn attachment plus four practices that follow from it, and three disclosure positioning strategies with the criterion for choosing between them.