Write social posts and captions in your own voice
A week of captions for one small clinic, on two platforms, drafted from five things you already wrote. Hooks, hashtags and the edit pass.
The Sunday night job
Take a clinic we will call Kestrel Physio: two therapists and a receptionist in a market town. There is an Instagram account with about nine hundred followers and a LinkedIn page for the GPs and two local employers who refer people. Nobody there is a social media manager. The posting gets done on Sunday evening, badly, by whoever remembers.
The week ahead is ordinary. A Reformer Pilates class starts on Thursday. They are closed for the bank holiday. They keep meaning to post about desk setups, because half the caseload is necks and shoulders. And there is a photo of the redecorated treatment room. Four things, two platforms, eight pieces of writing. On a good week they manage three.
That is the job here: eight captions by Sunday night that sound like the people whose names are on the door.
Paste five things you already wrote
Generic caption generators fail at one specific thing. Ask for "an Instagram caption about a new Pilates class" and you get something that could belong to any clinic in the country: an emoji, a rhetorical question, the word "journey".
The fix takes two minutes and it is the only part that really matters. Before you ask for a caption, paste in five things you have actually written. Pick ordinary ones: two captions from posts that went well, a message about where to park, the paragraph on your about page, the email confirming an appointment. Then say that this is how you write and every draft should start from that register.
Five samples work because they give the model range instead of a rule. Words like "friendly but professional" describe a voice from the outside. Samples show it from the inside, including the things you would never think to specify: that you write "we" and never "I", that you avoid exclamation marks, that you say patients rather than clients.
One brief, three angles
Then write the brief. One paragraph, once, covering all four items: what is happening, when, who it is for, what you want somebody to do about it. Facts only. Leave the tone out, because the samples are carrying that.
Now the move that saves the evening. Ask for three angles on the same item rather than three drafts of one angle. For the Pilates class: one that opens on the problem, a back that gives out by Wednesday; one on the thing itself, six people and one instructor on a Thursday morning; one on a person, the patient who could not reach her top shelf in March and can now. Same facts, three entry points. You will know which one is yours in seconds, a far faster decision than choosing between three versions of one sentence.
If all three come back sounding like each other, switch model on your next message instead of rewriting the prompt. Switching continues the same thread, so your samples and brief are still in play and only the answer comes from somewhere else. As of 19 October 2026 the picker holds 59 models from 17 providers, 22 of them on the free plan.
The first line carries the post
Ask for the first line on its own, six of them, for the post you care most about. Six is enough to include two you would never write, which shows you where the edge of your own register sits.
Our view is that the opening line does most of the work in a caption, because it is what somebody reads before deciding whether to read on. Reject the question that only pretends to be a question ("Ever wonder why your neck aches by Thursday?") and the statistic you have not checked. Keep the specific one. "Could not reach her top shelf in March" beats "improve your mobility", because one of those happened to somebody.
Hashtags, and how many survive
Ask for hashtags and you get a long block of them, most being the words in your post with the spaces taken out. Keep the ones you would type into a search box yourself. For this clinic that is the town, the discipline, and sometimes the class. Cut the aspirational ones, anything so broad that your post disappears inside it, and every single-word emotion.
After the first week, stop spending a model call here: write down the five you use and add one for the post.
Then take the tells out
Good samples get you most of the way, and drafts still arrive with habits that read as machine-written. The consistent ones in our own drafting: the balanced sentence where both halves are weighed against each other; the closing line that restates the post instead of asking for anything; "dive in", "unlock", "elevate", "game-changer"; and the rhetorical question in first position, the most reliable giveaway in a caption.
Then there is invented precision. A draft will cheerfully write "most desk workers" or "studies show". If you did not know the figure before you asked, it does not go in the post. In a clinic that matters more than tone, because a claim about what treatment achieves has to be yours.
The pass takes about a minute per caption. Read it aloud. Delete the first sentence if the second is stronger, which it often is. Swap one general noun for a specific one. A caption that arrived as "Ready to transform your core strength? Our brand new Reformer Pilates class is here!" comes out as something like this:
Six spots, Thursday mornings, 7.15. Reformer Pilates with Anna, who will not let you cheat the last five reps. It is for the people we see every week with backs that give out by Wednesday. Book at the desk or give us a ring.
Same facts. Nobody else could have posted it.
The LinkedIn version is then the quick one. Same brief, same thread, one sentence of new instruction: the reader is a GP or an HR lead, and the ask is a referral rather than a booking. Working both platforms in one thread is what makes the second caption cost a sentence instead of an evening.
When there is nothing to photograph
Three of Kestrel's four posts have an obvious picture: the room, the class, the closed sign. The desk-setup post has nothing, and photographing your own reception desk to illustrate a bad one is worse than posting no image.
So generate one. There are seven image models behind this, and on the web you will not see a button for them. Intent Detection is on by default for every account, so the request comes out of the message itself: ask in the chat for a simple illustration of a badly set up desk, and that is what answers. The toggle that brings the buttons back sits under Settings, then Tool Preferences. Each turn gives you one image, so three versions means three messages rather than a grid.
On the free plan you get two generations every 30 days, they run on Google Nano Banana, and they live in the web app, because on mobile the "Create image" control is a subscriber feature. Two a month is enough to find out whether generated images suit your feed, and for many small local accounts they do not: a wonky phone photo of the real treatment room usually beats a clean illustration of a room that does not exist. Save the generated image for the post where the alternative is a wall of text.
Keeping the voice past week one
None of this should be retyped every Sunday. On a paid plan, memories are where it belongs: tell the chat to remember how you write and it saves a short note, and Settings, then Memories, lists them and lets you write one by hand. Pro holds 200 and Pro+ has no cap. The free plan has none at all, worth knowing before you build a weekly routine on them.
Two things sit on every plan. Personal preferences and Traits, under Settings, then Account, apply to every new conversation, and the preferences box takes 5,000 characters, room for your five samples. Folders, 25 on the free plan, keep this month's social drafting in one place rather than scattered through your history.
One thing to know before you paste a year of your own writing into anything. We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker.
Where we would stop
The honest limit is that this writes the eight captions and it does not make the account worth following. Kestrel's best post of the year was a badly lit photo of a patient's dog asleep on the treatment couch, captioned in nine words. No brief produced that and no model would have proposed it.
What it does remove is the reason the other five posts never got written, because editing a few drafts is easier than starting from an empty page on Sunday evening. So our advice is narrow: use it for the ordinary eight, do the editing yourself, and keep the dog photo for when the dog turns up. If the memories part is what you are after, see the plans. The rest works on the free plan, and that is where we would start.