20 Gemini prompts for social media posts
Gemini has three things the other models do not: saved Gems, live search grounding and a spreadsheet at the end of it. These prompts use all three, and strip the formatting it adds by default.
The reason to prompt Gemini differently is not that it writes differently. It is that it sits next to three things the others do not have in the same place: a saved Gem you can hand your whole business to, a live search it can consult before it writes, and a spreadsheet you can drop the result into. Used as a plain chat window it is a chat window. Used properly it is closer to a small content desk.
Our articles are still written by humans!
Get human written articles in your Google feed.
It also has one habit that ruins captions, which is that it formats everything. Headings, bullets, bold, an emoji at the front of each line. That is right for a briefing document and wrong for a post, and it comes back unless you ban it somewhere permanent.
Build the Gem first
A Gem is a saved set of instructions plus files you upload as knowledge, described in Google's own help pages. It is the same idea as a project in Claude or a custom GPT, and it exists so you stop pasting your brief into every conversation. Google's own tips page is worth two minutes before you start.
1. Draft the Gem instructions
I am going to build a Gem that writes social posts for my business. First, interview me. Ask me up to eight questions, one at a time, that you actually need answered to write like us rather than like any business in our industry. Do not ask about brand values or mission. Ask about what we sell, who buys it, what we charge, what we refuse to do, and how we talk to customers in person.
2. Turn the answers into the Gem
Now write those answers as instructions for a Gem, under 350 words, written in the second person to yourself. Include a standing formatting rule: plain prose only, no headings, no bullet points, no bold text, no emoji unless I ask, and no introduction or sign-off around the output. Include a list of words we never use. End with the rule that if you do not know something about us you ask instead of inventing it.
Paste that into the Gem instructions and upload two files as knowledge: fifteen captions you have actually posted, and your price list or services page. Examples of your own writing do more than any adjective you could choose, which is the argument made at length in how AI learns your brand voice.
3. Try to break it before you trust it
Write me a caption about something we do not offer. Then write one that quotes a price I never gave you. If you correctly refused either of those, tell me which instruction stopped you. If you did not refuse, write the instruction that should have.
That third prompt is the one people skip, and it is the one that stops a confident invented claim reaching your feed six weeks from now. The failure is never announced when it happens.
Prompts that use the search
This is the part that genuinely differs from a model working from memory alone. Anything time-bound, local or seasonal is worth grounding, and every one of these prompts should end with you checking the answer.
4. What is actually happening near you
Search for what is happening in [town] over the next six weeks: events, school terms, markets, sport fixtures, roadworks. List only things you can find a source for, with the link next to each. Then tell me which three are relevant to a business like ours and which are not, and be honest when the answer is none of them.
5. The seasonal calendar for your actual trade
For [trade], what predictably changes month by month in demand, in what customers worry about, and in what we should be reminding them to do? Base it on searchable sources rather than on general marketing advice, and cite them. Where a month has nothing real in it, write nothing for that month rather than inventing an angle.
6. The questions people are asking, in their words
Search for how people actually phrase questions about [subject]: the clumsy, specific, real ones, not the tidy ones. Give me 15, in the wording you found them in. Do not clean them up. Then tell me which five we are genuinely qualified to answer.
7. Check a claim before it goes in a post
This sentence is going in a post: [paste it]. Search for whether it is currently true, and give me the source with its date. If the sources disagree, say so instead of choosing. If the only support is a marketing blog quoting another marketing blog, tell me that plainly and I will drop the claim.
That last prompt earns its keep more often than it should have to. Recycled statistics with no traceable origin are the standard currency of social media advice, and putting one in a post is a small, permanent liability. The same instinct, applied to your own website, is what actually gets you quoted by assistants, which is the subject of how to get your business to show up in ChatGPT.
Prompts for the posts themselves
8. One idea, three lengths
Write this post at three lengths: 25 words, 50 words and 120 words. Not the same sentences trimmed. Each version has to decide for itself what survives. Count the words and show the count. Plain prose, no formatting, output only.
9. The first line on its own
Write only the first line of this post, 10 times, maximum 9 words each. No questions, no greetings, no announcement that we are excited. Half should contain a number or a name. Then mark the ones that would be equally true of a competitor and delete them.
10. Caption for a photo you already have
Here is what is in the photo: [describe it plainly]. Write the caption. Under 45 words, one idea, no emoji, no question at the end. Do not describe the photo. Say the thing the photo cannot say, such as how long it took, what it cost, or what went wrong first.
11. The offer post that does not sound like an advert
Write a post about [offer]. Rules: state the price in the first line, name who should not buy it, and finish with one plain instruction. You may not use the words exclusive, unlock, elevate, seamless, or the phrase we are excited to announce. Under 60 words.
12. Same post, four channels, four readers
Rewrite this for Instagram, Facebook, LinkedIn and a Google Business post. Do not change the length and call it done. Instagram is for people who already follow us, Facebook is for the town, LinkedIn is for people in our trade, and Google Business is for somebody deciding right now. Keep every fact identical across the four.
The Google Business version is the one small businesses skip and the one with the shortest distance to money, because the person reading it is already looking for you.
Prompts for the month, in a spreadsheet
This is where Gemini being part of Google actually helps. Get the plan out of the chat and into a sheet, because a plan you cannot sort is a plan you will rebuild from scratch next month.
13. A month as a table you can paste into Sheets
Build a four week posting plan as a table with these columns and no others: date, channel, post type, subject in six words, first line, what the picture shows, and who this post is for. Twelve rows. Use only ideas drawn from what I have told you about this business. Leave a cell empty rather than filling it with something plausible.
14. Rebalance it honestly
Read that plan as a follower would over four weeks. How many of the twelve are asking them for something? How many require a photograph I do not currently have? Rewrite only the rows that fail those two tests, and tell me which rows you changed.
15. Turn one row back into a post
Take row 7 of the plan and write it out in full: caption, first line, hashtags separately, and one sentence describing the photo I need to take. Nothing else. Then stop and wait for me to say next.
How many rows should be in that table is its own argument, and the honest answer is fewer than most advice suggests. We worked through it in how often a small business should post.
Prompts for the picture
Gemini will generate an image, which is a genuine advantage over a text-only assistant and also the fastest way to make your feed look like everyone else's. Our rule is narrow and it has held up: generate the abstract thing, photograph the real thing.
16. A brief for a photo you take yourself
For this post, describe the single photograph I should take: what is in frame, where the phone goes, where the light comes from, and what to move out of shot first. It has to be takeable today, in a place I already am, with nothing I have not mentioned owning. If this post really needs a photo I cannot take, say so.
17. Generate only the abstract slide
Generate a plain background graphic for a quote slide: no people, no premises, no products, no text, no logos. Muted colours that would sit behind dark type. Square. It has to look like a background, not like a picture competing with the words.
18. Alt text that is actually useful
Write alt text for this image in one sentence, describing what is there rather than what it means. No phrases like image of. Under 20 words.
The reason for prompt 17's list of bans is simple. A generated photograph of a workshop that is not your workshop is a small lie told to the people most likely to notice, and it costs more than the ten minutes it saved.
Prompts for fixing the output
19. Strip the formatting
Give me that again as plain prose. No headings, no bullet points, no bold, no emoji, no numbered list, no explanation of what you did. If it was three bullets, make it three sentences.
20. Take out the sentences that give it away
Rewrite this so no sentence follows a template. Remove: in today's fast paced world, whether you are X or Y, it is not just A it is B, and any three item list where the third item is the emotional one. Replace each with something specific to us, or delete it. Show me the list of what you removed.
That second one is the whole of why AI posts sound like AI, compressed into a prompt you can paste. It is worth running on your own writing occasionally too, which is a humbling experience.
What the Gem still cannot do
It cannot know what happened in your business this week. Everything above is a way of shaping material, and the material comes from you writing down five lines about your own week, which takes two minutes and is the step almost everyone drops first.
It also does not survive contact with a busy month. The Gem does not get tired, but the person opening it does, and the honest comparison of doing this by hand against paying for something that runs without you is in is AI social media management worth it for a small business. What you get for nothing, and exactly where the ceiling sits, is in free AI social media tools.
Frequently asked questions
Can Gemini write social media posts?
Yes, and it is at its best when you build a Gem first rather than prompting from an empty chat. A Gem holds your business profile, your rules and examples of your own captions, so each later prompt can be a single line. The one thing to set immediately is a formatting rule, because Gemini's default output is headings and bullet points, which is wrong for a caption.
What is a Gem and do I need one for social media?
A Gem is a saved set of instructions plus files you upload as knowledge, which Gemini then applies to every chat inside it. For social media it is worth ten minutes because it removes the biggest cause of generic output, which is that you paste a shorter and shorter brief each week until there is nothing of your business left in the prompt.
How do I stop Gemini using bullet points and emoji in captions?
Put the rule in the Gem instructions rather than in the message: plain prose only, no headings, no bullets, no bold, no emoji unless asked, and no introduction or sign-off. Instructions given in a single message tend to hold for a few turns and then drift back, while instructions in the Gem apply every time.
Is Gemini better than ChatGPT for social media?
It is different rather than better. Gemini's advantages are the saved Gem, searching before it writes, and getting a plan out as a table you can paste into Sheets. Its disadvantage is a formatting habit you have to ban explicitly. The gap between models is far smaller than the gap between a prompt containing your actual business and one that does not.
Can Gemini make images for my posts?
It can, and the useful limit is to generate abstract things only: plain backgrounds, texture, a colour field behind a quote. Anything a customer could recognise, meaning your premises, your product or your team, should be a photograph you took. A generated version of a real place is noticed by exactly the people you least want to lose.
Can Gemini schedule my posts?
No. It writes and plans, and something else publishes. The usual arrangement is a plan built in Gemini, dropped into a sheet, then loaded into a scheduler. The tools that do the publishing half, and the ones that try to do both, are compared in our roundup of AI social media tools.
Should I trust what Gemini finds when it searches?
Treat it as current rather than as correct. Grounding a claim in a search result means it came from somewhere recent, not that the somewhere was right. For anything you would be embarrassed to have wrong in public, ask for the source and its date and open the link yourself before it goes into a post.
How we researched this
- These twenty are written for the Gem workflow described in the first section rather than for an empty chat window, because that is where the difference between Gemini and any other chat model actually sits. Each one assumes the business profile is already saved, which is why they are short. A prompt that only worked when a person quietly supplied the missing context was rewritten or dropped.
- The formatting prompts near the end are not padding. Gemini's house style is headings, bullet points and bold text, which is useful in a document and wrong in a caption, and it tends to return to that style unless told otherwise in the Gem rather than in the message. We wrote the instruction once so you can stop retyping it.
- Where a prompt asks Gemini to search before writing, the article says plainly that the result still has to be checked. Grounding a claim in a search result makes it likelier to be current, not likelier to be true, and a post that states a competitor's price wrongly is your problem rather than the model's.
- Feature descriptions were checked against Google's own documentation on 10 August 2026 and are linked below. Nothing here is a controlled comparison between models, because we have not run one, and the habits described are tendencies to check against what you get today rather than fixed properties. We sell a competing product, disclosed in the article where it becomes relevant.
Sources
