Give the AI your own voice as evidence, not just a topic. Paste 15 to 20 of your best past captions, have it describe your patterns, and save that description in a Project. For each new post, give it a raw note of what actually happened, ask for three options, pick one, and rewrite the first line yourself.
AI captions sound "off" for a simple reason. Asked for an Instagram caption with no examples, a model writes in the platform's average voice: sparkle emoji, three exclamation marks, "Swipe to see more!" That voice comes from millions of other accounts. The only way to pull it towards yours is to show it how you write, then to feed it the specific details only you know.
Build a voice file from your last 30 captions
Consider a church office that posts three times a week about a toddler group, community lunches, concerts, bell-ringing practice and the food bank collection. The administrator writes most captions; two volunteers help. Their best-performing posts are warm, plain, lightly funny and welcoming to people who don't come to services. That's the voice to capture.
Scroll back through your last 30 captions and copy the 15 to 20 you're proudest of, or that did best, into one document. Leave out anything written in a rush or by someone else. Then ask the AI to analyse, not write:
Below are 18 Instagram captions I wrote. Don't write anything new.
Describe my writing so another writer could copy it:
- typical caption length and sentence length
- how I open (first line patterns) and how I close
- emoji: which, how many, where
- punctuation habits (exclamation marks, dashes, questions)
- words and phrases I use often
- words I never use that most Instagram accounts do
- tone in three adjectives, with an example line for each
Quote my captions as evidence for every point.
[paste captions]An illustrative extract of the result:
Length: 40-90 words; short sentences, often under 10 words.
Openings: usually a plain fact or small scene ("The urn is on.",
"Forty-two jars of jam so far."). Never a question.
Closings: practical detail last (time, place, what to bring).
Emoji: 0-1 per caption, usually none; occasionally one at the end.
Exclamation marks: rare (3 in 18 captions).
Often used: "all welcome", "no need to book", "bring a friend"
Never used: "amazing", "blessed", "don't miss out", "link in bio!"
Tone: warm ("Come as you are, stay as long as you like"),
dry ("The bells are louder than they sound from the car park"),
practical ("Step-free entry by the side door").Check it against your captions before saving. In this run the model claimed the captions "frequently use humour", but only two of the 18 did. That was corrected to "occasional dry humour, one line at most", because an AI told you're frequently funny will try to be funny in every caption. Save the corrected description as the voice file in a ChatGPT or Claude Project, or a Gemini Gem (becoming a skill from November 2026), with your 18 examples attached. For a fuller version that covers your website and emails too, see building a brand voice guide AI can follow.
Give the AI a raw note, not a topic
The second half of the fix is the input for each post. A topic gives the model nothing to work with except its defaults. A raw note gives it your details.
Topic (what most people type): Write a caption about our community lunch.
Raw note (what works): Community lunch last Sat. 38 came, most ever. Soup was leek and potato, volunteers from the toddler group made the cakes. Two new families came because they saw last month's post. Next one Sat 15 Nov 12.30, free, no need to book. Photo: long table, people's backs, cakes in front.
The raw note takes 60 seconds to type on your phone right after the event. It contains a number, two concrete details, a small story (the families who came because of a post) and the next date. Every one of those is something the AI can't invent correctly, and every one makes the caption sound like a person who was there.
The caption prompt, and three options to choose from
With the voice file saved in the Project, the prompt for each post can stay short:
Using my voice file, write 3 caption options from this note.
Option 1: lead with the number. Option 2: lead with a small detail.
Option 3: lead with the next date.
40-80 words each. Facts only from the note. Max 1 emoji.
End with the practical details for the next one.
Note: [paste raw note]An illustrative set of options:
- "Thirty-eight for lunch on Saturday, our biggest yet. Leek and potato soup, and cakes from the toddler group team, who clearly have hidden talents. Two families came because they saw last month's post, so this is us asking you to share this one. Next lunch: Sat 15 Nov, 12.30, free, no need to book."
- "The cakes went first. They always do when the toddler group volunteers bake. Saturday's community lunch had 38 people round one long table, the most we've had. Next one is Sat 15 Nov at 12.30. Free, all welcome, no need to book."
- "Pencil in Saturday 15 November. Our community lunch is free, starts at 12.30, and needs no booking. Last time 38 of you came, which is a record, and the cakes vanished in about ten minutes."
Option 1 is the strongest because it asks readers to share, and gives them a reason (the two families). The administrator took it, but changed "clearly have hidden talents" to "who are now in charge of cakes permanently", which sounded more like the parish. That edit took twenty seconds, and it's the sort of change that keeps captions sounding human: small, specific, and in your own words.
When the draft gets a fact wrong
Even with a raw note, check the practical details line by line, because that's where small errors hide. In one illustrative run, option 3 above first read "Our community lunch is free, starts at noon, and happens every Saturday". The note said 12.30 and the lunch is monthly. Nobody would spot "every Saturday" as an AI error, but a family turning up the following week to an empty hall would remember it. The fix is a habit: before posting, read only the date, time, place and price against your note, separately from reading for voice.
First line, hashtags, alt text and search
A few platform details decide whether the caption gets read at all:
- The first line is the caption for most people. Instagram shows only roughly the first 125 characters before "more", and many people never tap. Write that line yourself, or at least check it stands alone. "Thirty-eight for lunch on Saturday, our biggest yet." works on its own; "So, where do we start?" doesn't.
- Five hashtags at most. Instagram capped posts and Reels at five hashtags in December 2025, down from 30. Pick a few specific ones (your town's community tag, an event tag) and skip the generic ones. Tell the AI the limit, because many models still suggest 20.
- Plain words help people find you. Since July 2025, public posts from professional accounts can appear in Google and Bing results, and captions do a lot of that work. "Free community lunch, church hall, Saturday 12.30" is searchable; "Good times with good people" isn't. You can check the setting for search-engine visibility in your account's privacy settings.
- Alt text. Add a short description of the photo for people using screen readers. AI drafts this well if you give it the photo and ask for one plain sentence, but check it: "people eating at a long table in a church hall, cakes in the foreground". Generating and checking AI alt text covers the details.
Reels and carousels need different captions
A photo caption carries the whole post. A Reel's caption doesn't, because people are watching, not reading, so the words on screen do most of the work and the caption can be one or two lines plus the practical details. A carousel is the reverse: the slides tell the story and the caption's job is to make people swipe and then act. Tell the AI which format it's writing for. An illustrative pair for the same event, the harvest jam stall:
Reel caption: "Forty-two jars, sold out by eleven. Next stall: Sat 8 Nov, 10am, church porch."
Carousel caption: "Where 42 jars of jam came from, in five slides: the plums, the volunteers, the labelling table, the queue, and the empty stall at eleven o'clock. Next stall Sat 8 Nov, 10am, church porch. Bring cash; we can't take cards yet."
The read-aloud edit: would you say it at the door?
Before posting, read the caption aloud and ask whether you'd say it to someone arriving at the door. If a phrase would sound odd spoken, it reads as AI. The usual culprits, and what the church office replaces them with:
| AI phrase | Why it grates | What we'd say |
|---|---|---|
| "Join us for a heartwarming afternoon of fellowship!" | Nobody talks like this | "Come for soup and stay for cake." |
| "Don't miss out on this amazing event!" | Pressure, and "amazing" means nothing | "There's room for everyone." |
| "We're so blessed to have such an incredible community ✨🙏" | Gushing; emoji stack | "Thank you to everyone who baked." |
| "Swipe to see more! 👉" | Stock instruction | Let the first photo make people swipe |
| "Whether you're a regular or new to our church…" | Template opener | "First time? Use the side door; it's step-free." |
Keep a running list of phrases you've cut and add it to the voice file as "never use". After a month, the drafts arrive needing far less editing. A written list of words your brand avoids is the same idea, formalised.
Different posts need different caption shapes
One voice doesn't mean one template. The church office found five types of post came up every month, each with its own shape:
| Post type | Shape | Example first line |
|---|---|---|
| Event coming up | What, when, who it's for, practical details last | "Bell-ringing practice is open to visitors on Thursday." |
| Event just happened | A number or detail, a thank-you, the next date | "Forty-two jars of jam, and the stall sold out by eleven." |
| Appeal | The need in one line, exactly what to give, where and by when | "The food bank needs tinned fish and long-life milk this week." |
| Behind the scenes | A small scene, a person (with permission), no ask | "The urn goes on at 9.15. By 9.30, the first toddler has found the biscuits." |
| Notice | The change, the reason, the alternative | "No toddler group on 29 Oct: the hall floor is being sealed." |
Add these shapes to the Project, and start each prompt with the post type ("Appeal post, from this note: …"). The appeal shape matters most, because appeals written in the AI's default voice tend to sound guilt-laden, which is the opposite of what works. Charity appeals and social posts that keep your voice goes deeper on that.
The same method gives a very different voice elsewhere
None of this makes captions sound churchy; it makes them sound like whoever's posts you feed in. A boutique hotel running the same voice analysis on its best 20 captions got a completely different file: 20 to 40 words, a sensory first line ("Log fire's lit."), no hashtags beyond its own, a light, confident tone, and never the words "luxury" or "nestled". Its raw notes read differently too: "First frost on the lawn, breakfast in the conservatory, the kitchen's doing the pear tart again". Same process, different evidence, different voice. That's the whole point: the AI doesn't have a good voice to give you, but it copies yours well once it can see it.
When three people write the captions
Consistency breaks when helpers use their own AI accounts with no voice file. In the church office's case, a volunteer covering during the administrator's holiday used a free AI app on her phone. Within a week, the feed had "✨ Exciting news! ✨" twice and a caption that called the food bank "our amazing initiative". Nothing was wrong, exactly; it just didn't sound like the church, and two regulars commented on it.
The fix was organisational rather than technical:
- One shared Project holding the voice file, the post shapes and the banned phrases. ChatGPT and Claude both offer shareable Projects on their business plans; on individual plans, keep the voice file in a shared document that each person pastes in.
- A one-line raw-note template pinned in the volunteers' group chat: "What happened / a number / one detail / next date / photo".
- A second pair of eyes for anything that isn't a routine event post: appeals, anything about a person, anything with children in the photo (and those only with parents' written consent).
Run captions through a short pre-publish checklist for AI-drafted posts until the habit sticks. For more ways to keep prompts consistent across people, see writing AI prompts that sound like your business.
Replies to comments, in the same voice
Comments are part of the voice too, and they're where an AI-sounding reply is most obvious. Draft replies with AI if it helps, but never auto-post them. A useful prompt, run with the voice file in the Project: "Draft a reply under 25 words to this comment, in my voice. If it's a question, answer it from the note; if you don't know, say we'll check." For "Is the lunch suitable for someone in a wheelchair?", the right reply names the step-free side door and the accessible toilet, which the model only knows if they're in your voice file or notes. Add those standing facts (access, parking, what's free) to the Project once, and every reply draft gets them right.
Checking whether your captions are working
Likes are a poor guide for a small account. Once a month, look at four things for each post: saves (people keeping the details), shares (people passing it on), comments and messages (people starting a conversation), and the real-world result where you can count it: how many came, how many items arrived, how many enquiries.
Paste a month's captions with those numbers into your AI assistant and ask it to compare the top five with the bottom five: what the top five did in their first lines, their length and their calls to action. In an illustrative review, the church office's top posts all opened with a number or a small scene, and the bottom ones opened with "We're delighted to announce". That became a rule in the voice file.
Finally, the time saved. Writing captions from scratch, the administrator was spending about 20 minutes per caption, three a week: an hour weekly. With the voice file, raw notes and three-option prompt, it's closer to six or seven minutes each, including the edit: about 20 minutes weekly. Over a year that's more than 30 hours back, and the captions sound more like the church than the rushed ones did.
Further reads
- How to Plan a Month of Marketing Content With AI in One Hour — Plan the month's posts before writing the captions.
- How to Create On-Brand Social Media Graphics With AI — Make the images match the voice of the captions.
- How to Turn Instagram Comments Into Leads With Keyword Auto-DMs — Turn caption calls to action into enquiries.
- Why Your AI Marketing Copy Sounds Generic (and How to Fix It) — The underlying reasons AI copy sounds the same.
- Should a Small Business Use AI to Write Its Social Media Posts? — Decide how much of your posting AI should do.
- How to Write LinkedIn Posts With AI Without Sounding Like AI — The same voice problem on a very different platform.
- How to Write Menu Descriptions With AI That Still Sound Like You — A five-stage method for AI-drafted menu descriptions that keep your kitchen's voice, with a voice card template, a drafting prompt and a before-and-after dish.
- AI Prompts for Bakers: Pricing, Captions and Customer Replies — Twelve copy-ready prompts for bakers covering costing and prices, social captions, and customer replies, each with what to check before you use the answer.
- How Photographers Can Fill Mini-Session Slots With AI Marketing — A three-week, waitlist-first campaign for mini sessions: slot maths, past-client segments, copy prompts and honest ways to fill the last slots.
- How to Plan a Month of Barbershop Instagram Content With AI — A one-hour planning session, a sample four-week grid, a shot list for filming the lot in one afternoon, and captions that sound like your shop.
- How to Create Outfit and Styling Posts for a Boutique With AI — Turn a boutique's stock list into weekly outfit carousels: AI plans combinations and writes the copy, while the photos stay real and the sizes stay honest.
- How Food Trucks Can Use AI to Plan Stock, Pitches and Posts — A trading log, a weekly planning routine and copyable prompts that help a food truck plan prep, choose pitches and post its schedule with AI.
- How Florists Can Use AI for Orders, Reviews and Social Posts — Worked examples for a flower shop: a rushed phone order turned into a docket, a substitution message, a year of reviews analysed, and accurate captions.
- Using AI in a Yoga or Pilates Studio While Keeping Your Voice — Why AI makes every studio sound the same, and the voice sheet, before-and-after edits and teacher-bio method that keep yours recognisably yours.
- AI Mistakes That Put Couples Off Your Wedding Business — Eight ways AI makes a wedding supplier look careless or fake to couples, each with an example from real-looking enquiries and the fix.
- How Agencies Keep Brand Voice Consistent in AI-Written Content — A filled-in voice card, two client voices side by side, a sample-led prompt and a five-line scoring rubric for keeping AI copy on-brand.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Instagram announcements on the five-hashtag limit (December 2025) and search-engine indexing of professional accounts' public posts (July 2025), as reported by Meta and industry press; general writing practice.