Skip to content
The Drawn Ledgerby Vivian 18+

AI studioChapter 2

Captions with AI

A model can draft captions, post text and replies in your persona’s voice once you’ve shown it that voice. Then you edit, so a person can hear a person in it.

Entry
A2
Category
AI studio
Checked
Reading
6 min
Sections
7

Evidence

VerifiedChecked against the original source.Vivian’s field notesFirst-hand experience, not a study.UnverifiedPlausible, not proven yet.

In 30 seconds

  • Give the model a short voice sheet and three real captions before you ask for anything.
  • Ask for several drafts, keep one, and rewrite it in your own words.
  • Cut what gives a machine away: ad words, stray hashtags, every line built the same way.
  • Suggestive captions are fine on most platforms. Explicit content and selling for sends are where the rules bite.
  • A sub’s name, messages and amounts never go into a chat with a model.

You will learn

  • What goes on a voice sheet and how to use it.
  • What a weak and a good request to a model look like.
  • How to edit a draft so it sounds human.
  • What the providers’ and the platforms’ rules mean for teasing captions.
  • How to get help with a reply without handing over a sub’s private data.

A model can write a hundred captions a minute. None of them is yours until you’ve cut it down and said it your way.

Vivian

Show it the voice first

With no instructions, a model writes like an ad. So before you ask for a caption, give it a voice sheet: a few lines on who the persona is and how she talks. In my experience, ten lines work better than two pages, because the model follows a short sheet closely and a long one loosely. Vivian’s field notesFirst-hand experience, not a study.

A voice sheet in five parts

  1. Who she is, in one line: her name, her look, the kind of dom she plays.

  2. How she talks: how long her sentences are, words she uses, words she never uses.

  3. How she treats the reader: teasing, cold, warm or mocking.

  4. Three captions you’ve posted and liked, pasted as they are.

  5. What she never writes: explicit sex, real names, promises about money.

The examples do more work than the description. A model picks up rhythm from three real captions far better than from words like “playful” or “dominant”. Swap the examples every few weeks, or every draft starts to sound like the same three posts. Vivian’s field notesFirst-hand experience, not a study. Most of the sheet can come straight from your persona file, as Build a persona explains.

Ask for ten, keep one

Ask for ten drafts at once, each under a set length, and pick the best. Choosing is faster than fixing one bad draft. Say what the post is for: a new picture, a wishlist item, a tool launch, a quiet day. Vivian’s field notesFirst-hand experience, not a study.

A weak request

Tap every message that shows a red flag. There are two.

Example conversation. Invented for this exercise.

0 of 2 found

A good request

Tap every message that shows a red flag. There are three.

Example conversation. Invented for this exercise.

0 of 3 found

Make it sound like you

Every draft gets a human pass before it goes out. Read it out loud, and if you wouldn’t say it, change it. Vivian’s field notesFirst-hand experience, not a study. Then check it still says one clear thing, as Posting and workflows explains.

What gives a machine away

  • Red flag 1Ad words: “exciting”, “amazing”, “don’t miss”, “check out”.
  • Red flag 2Hashtags and emoji you didn’t ask for.
  • Red flag 3Every line built the same way, often in threes.
  • Red flag 4A question tacked onto the end of every post.
  • Red flag 5A long word where she’d use a short one.
  • Red flag 6The same opener day after day.

Keep a note of the changes you make most often. After a month, add them to the voice sheet, so the model stops making the same mistakes. Vivian’s field notesFirst-hand experience, not a study.

OpenAI’s terms list presenting output as human-generated when it wasn’t among the things users may not do. VerifiedChecked against the original source.OpenAI Terms of Use (outside the US and EEA) (opens in a new tab) Accessed 7 November 2026. Read your own provider’s terms on this before you post drafts as they come.

What the providers allow

A teasing caption meets the providers’ rules on one side and the platforms’ rules on the other. Vivian’s field notesFirst-hand experience, not a study. Pick your AI has the full table of providers. For captions, a few lines matter most.

Z.ai’s terms forbid suggestive content as well as explicit content, so even a teasing caption is outside them. VerifiedChecked against the original source.Z.ai Terms of Use (opens in a new tab) Accessed 7 November 2026. Anthropic forbids content related to sexual fetishes or fantasies, for every Claude model. VerifiedChecked against the original source.Anthropic Usage Policy (opens in a new tab) Accessed 7 November 2026. Kimi’s consumer terms forbid content that promotes obscenity or pornography. VerifiedChecked against the original source.Kimi Terms of Service (international, consumer) (opens in a new tab) Accessed 7 November 2026. xAI’s consumer FAQ says Grok may answer with sexual situations, among other things. VerifiedChecked against the original source.xAI consumer FAQ (opens in a new tab) Accessed 7 November 2026. On xAI’s API, a developer can tell the model to decline sexual innuendo, and leaving that setting out keeps the default behaviour. VerifiedChecked against the original source.xAI API docs: Custom moderation controls (opens in a new tab) Accessed 7 November 2026. No policy read names findom, so whether a money-teasing caption counts as fetish content is the provider’s call. UnverifiedPlausible, not proven yet.

If a model won’t write the captions you need, use one whose rules allow them, or write that kind yourself. Prompting around a model’s rules can cost you the account. Vivian’s field notesFirst-hand experience, not a study.

What platforms flag

Suggestive and explicit are two different lines, and most platforms draw theirs between them. Vivian’s field notesFirst-hand experience, not a study. Throne allows suggestive but not explicit content on a wishlist, and applies its rules to drawn art as it does to photos. VerifiedChecked against the original source.What’s allowed in my profile text and images? (opens in a new tab) Accessed 3 November 2026. X allows adult content that’s labelled and not prominently displayed, including anime and AI-generated art. VerifiedChecked against the original source.Adult Content Policy (opens in a new tab) Accessed 3 November 2026.

The warnings X describes go on photos and videos, either account-wide or one post at a time. VerifiedChecked against the original source.Your media settings (opens in a new tab) Accessed 3 November 2026. What X’s classifier reads in a caption’s text isn’t published. UnverifiedPlausible, not proven yet. The full method for marking posts is in Posting and workflows.

The words can still break a rule on their own. X’s adult content rules bar promoting or soliciting sexual services. VerifiedChecked against the original source.Adult Content Policy (opens in a new tab) Accessed 3 November 2026. Throne bars profile text that offers content or services in exchange for gifts. VerifiedChecked against the original source.What’s allowed in my profile text and images? (opens in a new tab) Accessed 3 November 2026. A caption that sells a picture for a send can fall under those lines, so read them before you write one. Vivian’s field notesFirst-hand experience, not a study.

Keep the sub out of it

Replies are where AI help gets risky, because the material belongs to someone else. A sub’s messages, name, handle, photos and amounts are his. Pasting them into a chat hands them to a company he never agreed to. Vivian’s field notesFirst-hand experience, not a study.

Help with a reply, without his data

  1. Describe the situation in your own words: “a sub asks for a bigger task than the one agreed”.

  2. Write “he” or “the sub” instead of his name or handle.

  3. Leave out amounts, or round them.

  4. Never paste screenshots of the chat.

  5. Ask for three short replies in the persona’s voice, then rewrite the one you pick.

Write prices, totals and dates yourself. A model can get a number wrong, and in a money message the mistake is yours. Vivian’s field notesFirst-hand experience, not a study.

Some replies shouldn’t be drafted by a model at all. If a sub says money is short, mentions debt or sounds unwell, stop the money part and answer him yourself, in plain words. Vivian’s field notesFirst-hand experience, not a study. See When a sub is in trouble.

What it costs

Captions don’t need the API. A chat plan is enough, and the providers below also have a free tier. Vivian’s field notesFirst-hand experience, not a study.

Scroll the table sideways

What it costs
PlanPrice read on 2026-11-07
ChatGPT Go and PlusShown in your local currency
Claude ProUSD 20 a month, billed monthly
Grok PlusShown in your local currency

Claude Pro costs USD 20 when billed monthly. VerifiedChecked against the original source.Claude plans and pricing (opens in a new tab) Accessed 7 November 2026. ChatGPT and Grok show their plans in the visitor’s own currency, so read the price on their pricing pages. UnverifiedPlausible, not proven yet. Prices depend on your country and change often. UnverifiedPlausible, not proven yet. If you build a tool that writes captions for you, the API prices are in Pick your AI.

A long message

A sub sends a long message about a task he finished, with his real first name in it, and asks what comes next. You want help with the answer.

Remember it

  1. Give the model a short voice sheet and three real captions first.
  2. Ask for ten drafts, keep one, then rewrite it.
  3. Cut ad words, stray hashtags and lines that all sound the same.
  4. Teasing is fine on most platforms. Explicit content and selling for sends are where the rules bite.
  5. A sub’s messages, name and amounts stay out of the chat.

Check it

Every claim with its evidence level and its sources.

  1. With no instructions, a model writes like an ad.

    Vivian’s field notesFirst-hand experience, not a study.

  2. So before you ask for a caption, give it a voice sheet: a few lines on who the persona is and how she talks.

    Vivian’s field notesFirst-hand experience, not a study.

  3. In my experience, ten lines work better than two pages, because the model follows a short sheet closely and a long one loosely.

    Vivian’s field notesFirst-hand experience, not a study.

  4. The examples do more work than the description.

    Vivian’s field notesFirst-hand experience, not a study.

  5. A model picks up rhythm from three real captions far better than from words like “playful” or “dominant”.

    Vivian’s field notesFirst-hand experience, not a study.

  6. Swap the examples every few weeks, or every draft starts to sound like the same three posts.

    Vivian’s field notesFirst-hand experience, not a study.

  7. Ask for ten drafts at once, each under a set length, and pick the best.

    Vivian’s field notesFirst-hand experience, not a study.

  8. Choosing is faster than fixing one bad draft.

    Vivian’s field notesFirst-hand experience, not a study.

  9. Say what the post is for: a new picture, a wishlist item, a tool launch, a quiet day.

    Vivian’s field notesFirst-hand experience, not a study.

  10. Every draft gets a human pass before it goes out.

    Vivian’s field notesFirst-hand experience, not a study.

  11. Read it out loud, and if you wouldn’t say it, change it.

    Vivian’s field notesFirst-hand experience, not a study.

  12. After a month, add them to the voice sheet, so the model stops making the same mistakes.

    Vivian’s field notesFirst-hand experience, not a study.

  13. OpenAI’s terms list presenting output as human-generated when it wasn’t among the things users may not do.

    VerifiedChecked against the original source.OpenAI Terms of Use (outside the US and EEA) (opens in a new tab) Accessed 7 November 2026.OpenAI Terms of Use (outside the US and EEA) (opens in a new tab)Accessed 7 November 2026

  14. A teasing caption meets the providers’ rules on one side and the platforms’ rules on the other.

    Vivian’s field notesFirst-hand experience, not a study.

  15. Z.ai’s terms forbid suggestive content as well as explicit content, so even a teasing caption is outside them.

    VerifiedChecked against the original source.Z.ai Terms of Use (opens in a new tab) Accessed 7 November 2026.Z.ai Terms of Use (opens in a new tab)Accessed 7 November 2026

  16. Anthropic forbids content related to sexual fetishes or fantasies, for every Claude model.

    VerifiedChecked against the original source.Anthropic Usage Policy (opens in a new tab) Accessed 7 November 2026.Anthropic Usage Policy (opens in a new tab)Accessed 7 November 2026

  17. Kimi’s consumer terms forbid content that promotes obscenity or pornography.

    VerifiedChecked against the original source.Kimi Terms of Service (international, consumer) (opens in a new tab) Accessed 7 November 2026.Kimi Terms of Service (international, consumer) (opens in a new tab)Accessed 7 November 2026

  18. XAI’s consumer FAQ says Grok may answer with sexual situations, among other things.

    VerifiedChecked against the original source.xAI consumer FAQ (opens in a new tab) Accessed 7 November 2026.xAI consumer FAQ (opens in a new tab)Accessed 7 November 2026

  19. On xAI’s API, a developer can tell the model to decline sexual innuendo, and leaving that setting out keeps the default behaviour.

    VerifiedChecked against the original source.xAI API docs: Custom moderation controls (opens in a new tab) Accessed 7 November 2026.xAI API docs: Custom moderation controls (opens in a new tab)Accessed 7 November 2026

  20. No policy read names findom, so whether a money-teasing caption counts as fetish content is the provider’s call.

    UnverifiedPlausible, not proven yet.

  21. If a model won’t write the captions you need, use one whose rules allow them, or write that kind yourself.

    Vivian’s field notesFirst-hand experience, not a study.

  22. Prompting around a model’s rules can cost you the account.

    Vivian’s field notesFirst-hand experience, not a study.

  23. Suggestive and explicit are two different lines, and most platforms draw theirs between them.

    Vivian’s field notesFirst-hand experience, not a study.

  24. Throne allows suggestive but not explicit content on a wishlist, and applies its rules to drawn art as it does to photos.

    VerifiedChecked against the original source.What’s allowed in my profile text and images? (opens in a new tab) Accessed 3 November 2026.What’s allowed in my profile text and images? (opens in a new tab)Archive (opens in a new tab)Accessed 3 November 2026

  25. X allows adult content that’s labelled and not prominently displayed, including anime and AI-generated art.

    VerifiedChecked against the original source.Adult Content Policy (opens in a new tab) Accessed 3 November 2026.Adult Content Policy (opens in a new tab)Archive (opens in a new tab)Accessed 3 November 2026

  26. The warnings X describes go on photos and videos, either account-wide or one post at a time.

    VerifiedChecked against the original source.Your media settings (opens in a new tab) Accessed 3 November 2026.Your media settings (opens in a new tab)Archive (opens in a new tab)Accessed 3 November 2026

  27. What X’s classifier reads in a caption’s text isn’t published.

    UnverifiedPlausible, not proven yet.

  28. X’s adult content rules bar promoting or soliciting sexual services.

    VerifiedChecked against the original source.Adult Content Policy (opens in a new tab) Accessed 3 November 2026.Adult Content Policy (opens in a new tab)Archive (opens in a new tab)Accessed 3 November 2026

  29. Throne bars profile text that offers content or services in exchange for gifts.

    VerifiedChecked against the original source.What’s allowed in my profile text and images? (opens in a new tab) Accessed 3 November 2026.What’s allowed in my profile text and images? (opens in a new tab)Archive (opens in a new tab)Accessed 3 November 2026

  30. A caption that sells a picture for a send can fall under those lines, so read them before you write one.

    Vivian’s field notesFirst-hand experience, not a study.

  31. Replies are where AI help gets risky, because the material belongs to someone else.

    Vivian’s field notesFirst-hand experience, not a study.

  32. Pasting them into a chat hands them to a company he never agreed to.

    Vivian’s field notesFirst-hand experience, not a study.

  33. A model can get a number wrong, and in a money message the mistake is yours.

    Vivian’s field notesFirst-hand experience, not a study.

  34. If a sub says money is short, mentions debt or sounds unwell, stop the money part and answer him yourself, in plain words.

    Vivian’s field notesFirst-hand experience, not a study.

  35. A chat plan is enough, and the providers below also have a free tier.

    Vivian’s field notesFirst-hand experience, not a study.

  36. Claude Pro costs USD 20 when billed monthly.

    VerifiedChecked against the original source.Claude plans and pricing (opens in a new tab) Accessed 7 November 2026.Claude plans and pricing (opens in a new tab)Accessed 7 November 2026

  37. ChatGPT and Grok show their plans in the visitor’s own currency, so read the price on their pricing pages.

    UnverifiedPlausible, not proven yet.

  38. Prices depend on your country and change often.

    UnverifiedPlausible, not proven yet.

Download