Blog
How to write good ChatGPT or Claude prompts for social media posts
A good social media prompt gives the model what a new person on your team would need to write that post: who the brand is and how it talks, who the post is for, what the post has to achieve, which network and format it’s going out on, what it can’t say, and three of your posts that worked. Then you ask for several versions with criteria for choosing between them, correct with specific instructions, and, before posting, check the facts, promises, images and tone yourself.
It works for ChatGPT, Claude or Gemini: the context, examples, role and structure follow what all three providers’ guides recommend; the versions with criteria and the final check are our own method. Below is the method applied to a sample brand, with four complete prompts for different tasks that you can adapt to yours.
What OpenAI, Anthropic and Google recommend
All three prompting guides are written for people using the models through the API, but what they ask for carries over to the chat. We read all three on November 17, 2026: Anthropic’s prompting best practices (the old “Be clear and direct”, “Use examples” and “Give Claude a role” pages now redirect there), OpenAI’s prompt engineering guide and Google’s prompt design strategies for Gemini.
| Technique | Anthropic (Claude) | OpenAI (ChatGPT) | Google (Gemini) |
|---|---|---|---|
| Clarity and context | Treat the model like a brilliant but new employee who lacks context on your norms and workflows. Test: if a colleague with no context would be confused by your prompt, the model will be too | GPT models perform best with explicit instructions, like a junior coworker; a reasoning model, like a senior coworker, can be given the goal and trusted to work out the details | Clear, specific instructions, and include the information the model needs instead of assuming it already has it |
| Examples | 3 to 5, close to your real use case, varied, and kept apart from the instructions | Show varied inputs with the output you want | Always include examples, in the same format; too many can make it overfit to them |
| Role | Giving it a role focuses its behavior and tone | An identity section: purpose, communication style and goals | The role can go in the system instructions or at the start of your message |
| Structure | XML tags to separate instructions, context and examples | Markdown headers and XML tags to mark sections | Tags or Markdown, using one format per prompt |
| Iterating | Before refining a prompt, define success criteria and a way to test against them | Tests that measure the prompt’s behavior as you iterate | It’s an iterative process: its templates are starting points |
Two details change what you do in the chat. Google says the role and the response format can go at the start of your message, so you don’t need API access to set them. And Google recommends splitting tasks up: one prompt per instruction, chained together when a task has several steps, with the output of one becoming the input of the next. That’s why there are four prompts below instead of one long one.
The context block, with a sample brand
Fictional case. Brasa Roasters is a brand made up for this article: it roasts specialty coffee and sells it through its online store and a café. Its people, products and posts are fictional too.
The context that doesn’t change from one post to the next gets written once and pasted at the start of the conversation. It has four parts:
- The brand and how it talks, with rules you can apply to a sentence (“short sentences, talk to one person”) rather than just adjectives (“friendly, authentic”). If you don’t have them yet, the voice and tone guidelines will help you write them.
- The audience: who reads, what they know and what they’re missing.
- The rules: what can’t be promised and what to do when a fact is missing. The instruction to write MISSING instead of making something up follows Google’s Gemini template, which asks the model to state that the information is not available when the answer isn’t in the context.
- Three posts that worked, with the criterion you used to pick them (most saves, most questions) and a note to use them for tone, not to copy sentences.
The other two pieces, the goal of the post and the network with its format, change with every request, so they go in each task’s prompt.
Brasa Roasters context block (fictional case)
You're the social media manager for Brasa Roasters. I'm going to ask you for several posts in this conversation: use this context for all of them unless a request says otherwise. <brand> Brasa Roasters roasts specialty coffee and sells it through its online store and its café. It talks like the barista serving you: short sentences, addressed to one person, and it explains a technical term the first time it uses it. One emoji per post at most. </brand> <audience> People who drink coffee at home every day and want it to taste better, without professional equipment. Most of them use a moka pot or a French press. </audience> <rules> Use only the facts I give you in each request. If you need a fact I didn't give you (a price, opening hours, a date), don't fill it in: write MISSING and what the fact is. Don't promise results, health benefits or shipping times. Write the way you'd talk to a customer at the café counter. </rules> <examples> These three posts got more saves and questions than the average for their format. Use them as a reference for tone and for how we explain things; don't repeat their sentences. <example> Carousel: Is your moka pot coffee coming out bitter? Here are the three things we check first when someone asks us that at the café. Swipe. </example> <example> Reel: This is what a roast sounds like from the inside. That cracking you hear halfway through is the signal we wait for to decide when to take it out. </example> <example> Photo: This season's coffee is here. You can taste it at the café before you take a bag home. Ask the barista why we picked it. </example> </examples>
The lowercase tags (<brand>, <audience>) are the structure Anthropic recommends; if you find it easier, plain-text headings like “BRAND:” and “AUDIENCE:” do the same job, as long as you stick to one format across the prompt, as Google asks. The three examples are different on purpose (a carousel, a Reel and a photo): Anthropic asks for varied examples so the model doesn’t pick up a pattern you didn’t intend.
Four prompts for different tasks
Each one is sent after the context block, in the same conversation. If you open a new conversation, paste the block first.
1. A Reel caption, with three versions and criteria
Instagram Reel caption (fictional case)
Task: write the caption for an Instagram Reel. The video: 30 seconds long. Valentina, our roaster, opens the first sack of this season's coffee, roasts it and tastes it. No voiceover, just music. Goal: get people who watch it to come and taste the coffee at the café this weekend. Confirmed facts: it's a medium roast; at the cupping we noted chocolate and orange; you can taste it for free at the café on Saturday and Sunday. I want: 1. Three versions that change only the first line: one that opens with the flavor, one with the process and one with the invitation. Keep the rest the same in all three. 2. Between 40 and 70 words each, ending with the invitation to the café. 3. Under each version, rate it with yes or no and one sentence per criterion: Does the first line make sense without watching the video? Does it use only the confirmed facts? Does the invitation say what to do and when? Does it sound like the examples? 4. At the end, tell me which one you'd post and why, and what fact you were missing.
What each part does:
- The video description. The model only knows what you tell it about the Reel. If you don’t say what it shows, you’ll get a caption that may have nothing to do with the footage.
- One variable between versions. If all three change the hook, the length and the ending at once, you won’t know what you liked about the one you pick.
- The written criteria. They’re the same ones you’ll use to decide. Anthropic asks you to define success criteria before refining a prompt; here the model uses them to check itself and you use them to compare.
- “What fact you were missing.” The café’s opening hours, for example. Better to find out now than to read a made-up one in the post.
2. A LinkedIn post for a different audience
LinkedIn post (fictional case)
Task: a LinkedIn post for the profile of Martín, founder of Brasa Roasters. For this post the audience changes: people who buy coffee for the office where they work, not people who drink coffee at home. The rest of the brand context stays the same. Confirmed facts: since March we've been delivering whole-bean coffee to offices in the city once a week. What people most often ask us to change is the grind, because many offices have drip coffee makers. We learned to ask which coffee maker they have before the first delivery. Goal: get interested offices to send Martín a direct message. Format: between 120 and 200 words, first person, paragraphs of one or two sentences. No hashtags in the body; three at most at the end. Don't name any offices or add numbers that aren't in the facts. Before you write, tell me in one line which part of the brand context doesn't apply to LinkedIn and how you'll adjust it. Then give me two versions: one that opens with the mistake we used to make and one that opens with the question we ask now.
When a task contradicts the context block (here, the audience), say so explicitly, as in the second line. Otherwise the model may try to satisfy both and writes for an audience that doesn’t exist. Asking it to say what it will adjust before writing shows you whether it understood the change.
3. Replies to the comments you get most often
Replies to frequent comments (fictional case)
Task: draft replies to the comments we get most often on our posts. We'll review them and save them as a starting point; each reply gets adjusted to the comment before it's sent. Frequent questions and the correct answer: 1. Do you ship? Yes, nationwide, through the online store. The delivery time depends on the destination and shows up before you pay. 2. Does it work in a moka pot? Yes. When you order, ask for a moka pot grind in the online store or at the café. 3. How long does it stay fresh? We recommend drinking it within a month of the roast date, which is printed on the bag. 4. Do you have decaf? Not right now. Rules: - Two versions per question: one to reply publicly, two sentences at most, and one to send by direct message if the person needs more detail. - Don't add information that isn't in the correct answer. - If someone complains about an order, a charge or a health issue, don't suggest a reply: write ESCALATE. Orders are handled by the online store team. - At the end, give me three questions that could come in and that this list doesn't cover, so I can decide the answer.
Here the model shapes answers you’ve already decided. The cases that need a person are ruled out in writing. The last instruction helps you find gaps in your list before they show up in a comment. When to reply publicly and when by direct message is covered in social customer service in comments and DMs.
4. Repurposing the LinkedIn post as a carousel
From LinkedIn to an Instagram carousel (fictional case)
Task: turn the LinkedIn post we approved (it's at the end) into an Instagram carousel for our usual audience: people who make coffee at home. What to keep: the idea of asking which coffee maker you have before choosing the grind. What to change: this audience isn't buying for an office. Remove the offices and the invitation to message Martín. Format: 6 slides, 20 words at most each. The first one asks the question and the last one says what to do: choose the grind when you buy. Give me: 1. The text for each slide. 2. Which photo you'd put on each slide, using only things a café can photograph: bags, grinders and coffee makers. 3. The post caption. At the end, check whether any sentence only makes sense on LinkedIn. Approved post: [paste the LinkedIn post you approved here]
This is the chaining Google describes: the approved output of prompt 2 is the input of prompt 4. Starting from the approved version, not the original idea, keeps sentences you already corrected from coming back. Deciding what to keep and what to change is your call; the model carries it out. To choose what to leave out when moving between formats, see how to repurpose content, and to order the slides, the carousel structure.
How to fix the result without starting over
If none of the versions work, don’t ask for “three more”: say what failed and why. Anthropic explains that giving the reason behind an instruction helps the model generalize from it, and, for formatting, recommends telling it what to do instead of what not to do.
Three more adjustments that often fix the problem:
- If it copies your examples almost sentence by sentence, swap one of them for a different post or remind it that they’re a tone reference. Google warns that with too many examples the model may overfit to them.
- If it ignores a rule, ask what it understood from that rule before rewriting it. Sometimes the problem is that the rule can be read two ways.
- When a prompt works, save it along with the context block. If you tweak it, change one thing at a time so you know what made it better. Google notes that the order of the content in the prompt can also change the response.
What to check every time before you post
The model works with what you gave it. What wasn’t in the prompt, and what changed since, is yours to check:
- Facts. Compare every price, date, opening hour, name and address against the source, not against your memory. Search the text for MISSING and for any number that didn’t come from your facts.
- Promises. Look for results, timelines or benefits the brand can’t deliver or back up. “The best coffee you’ll ever taste” is a promise, even if it sounds like style.
- Images and people. If someone appears in the photo or video, like Valentina in the Reel or a customer at the café, confirm they agreed to be in a post. If the image or the music isn’t yours, confirm you have permission to use it.
- Tone. Read the text without the brand name: if any competitor could sign it, swap one sentence for a fact of your own. It’s the five-minute review we suggest for any AI-assisted text.
- The moment. The model doesn’t know whether there’s an open complaint today, a change in opening hours or a news story that would make the post land badly. Only you know that.
What’s worth delegating to AI, and which decisions stay with a person, is laid out in what to delegate to AI in social media management.
From prompt to post, without copying and pasting
With this method, the chat gives you text and you copy it into each network. If you connect HeyMark to Claude or ChatGPT through MCP (in ChatGPT, with developer mode on), the same conversation can read your past posts and their metrics before suggesting anything (so the posts that worked come from your real data), leave the draft on your calendar, schedule it, and reply to comments and direct messages. HeyMark publishes to five networks: Instagram, TikTok, Facebook, LinkedIn and YouTube. Publishing now and replying wait until you approve the preview; scheduling runs right away, so check the date, time and text before you ask for it. The pre-posting checklist is still yours.
Plan your next post in HeyMark.
Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.