How to write AI Instagram captions that sound like you

Why AI captions sound generic

Without context, an AI writes the average of everything it has seen: an exclamation-heavy opener, three adjectives, a question and a stack of hashtags. That average belongs to nobody. The fix is not a cleverer prompt. It is better material: how you talk, what you know to be true and what you have already approved.

Write a one-page voice guide

Keep it short enough that you would actually read it. Each line should be something an editor could check.

• Three qualities, each with a boundary. "Warm, not gushing. Practical, not dry. Confident, never promising results."
• How you address customers: first name, formal or informal, singular or plural in languages where it matters.
• Words you use and words you never use, including industry jargon and phrases you dislike.
• Sentence length and rhythm. Short lines or flowing paragraphs.
• Emoji and punctuation habits: none, one at the end, or used as bullets.
• How you end: a question, a direct instruction or a sign-off.
• Hashtags: how many, and which ones you actually use.

Keep facts separate from style

Your voice guide says how to write. A separate fact sheet says what is true: product names, ingredients or materials, prices, opening hours, delivery areas, current offers and their end dates. Give the AI only the facts relevant to the post and tell it to use nothing else.

Update the fact sheet when the business changes. A beautiful caption cannot fix an old price.

A prompt structure that works

Whatever tool you use, give it the same building blocks every time. Consistency in the brief produces consistency in the voice.

• Goal of the post: attract, explain, prove or invite.
• Who it is for: new customers, regulars, a specific group.
• What the image or video shows, described plainly.
• The facts to use, pasted from your fact sheet.
• The call to action and where it leads.
• Length and format: "Under 60 words, two short paragraphs, one question at the end."
• Two or three approved captions as examples of the voice.

What AI should never invent

Treat any of the following in a draft as an error unless you supplied it.

• Prices, discounts, deadlines and stock levels.
• Reviews, testimonials, customer names or quotes.
• Results, statistics and before-and-after claims.
• Awards, certifications, "best in the city" claims and comparisons with competitors.
• Medical, financial or legal claims.
• Addresses, opening hours and event details.
• Partnerships or collaborations that are not confirmed.

Check every draft in under a minute

A quick, repeatable check catches most problems before approval. Read the draft once as your customer and once as your editor.

• Read it aloud. Would you say this at the counter or on the phone?
• Underline every fact and find it on your fact sheet. Delete anything you cannot find.
• Check the first line: it should make sense even if the caption is cut off.
• Check the call to action: one action, and the link or keyword is ready.
• Cut adjectives that any competitor could use about themselves.

Edit, then teach

Your edits are the most valuable voice data you have. Keep a list of approved captions, and for rejected drafts note why in one line: "too salesy", "we never say luxury", "wrong price". Add those notes to the voice guide. Within a few weeks, drafts need fewer edits. IGP one keeps this kind of brand memory for you: it learns from what you approve and correct, and uses your confirmed business facts in every draft.

If you write in more than one language, keep one voice guide per language. The personality stays the same, but greetings, sentence length and calls to action should sound native, not translated.

Checklist

1. Voice guide: three qualities with boundaries, words to use and avoid.
2. Fact sheet kept separately and updated when things change.
3. Five to ten real captions saved as examples.
4. Every prompt includes goal, audience, visual, facts, call to action and length.
5. Every draft checked for invented prices, results, reviews and claims.
6. Rejected drafts noted with a one-line reason.
7. One voice guide per language.