ChatGPT Prompts for Social Media Captions by Platform

Ask ChatGPT for “10 engaging Instagram captions” and it will hand you ten that open with an emoji, ask a rhetorical question, and end with “Link in bio!” They are not wrong. They are the average caption, and the average caption is what every feed is already full of.

A caption prompt has to know three things the generic request leaves out: where the platform cuts the text off, what the first line has to do before that cut, and what this specific post is for. I build all three into the prompt. Below is a base prompt, one prompt per platform, and the working prompts I use to turn one piece of content into a week of posts without the week reading like one post repeated.

What the prompt has to know

The fold is the constraint, not the limit. Every platform gives you more room than you should use and cuts the text off long before the limit. LinkedIn allows 3,000 characters in a post, but the feed shows a couple of lines before “see more”. TikTok’s publishing API caps a caption at 2,200 characters, and the app shows a single line before “more”. Instagram and X behave the same way: a short visible fragment, then a tap. The caption’s job is to earn that tap, so the first line has to work alone. Check each platform’s current limits before you paste them into the prompt, because they change.

The first line is the whole ad. Above the fold you have room for one sentence. That sentence has to be specific enough to stop a thumb. “We’re excited to announce” has never stopped a thumb.

One post, one job. A caption that tries to inform, entertain and sell in four lines does none of them. The prompt asks for the job up front so the model can cut everything that does not serve it.

Voice is set once, elsewhere. These prompts assume a voice block is already in place. If you do not have one, build a brand voice system prompt first. Without it, every prompt here produces the average caption in a slightly different shape.

The base prompt

Set this up once per brand, after your voice block.

You write social media captions for [brand: what it sells, to
whom, in one sentence].
Audience on social: [who follows us and why, one sentence].

Rules for every caption:
- The first line must work on its own, because the platform cuts
  the text off after it. It states something specific: a number,
  a claim, a situation the reader recognises. No questions, no
  emoji, no "we're excited".
- One job per caption. I will name the job. Cut anything that
  does not serve it.
- No hashtags inside sentences. Put them where I say, or omit.
- Never invent a statistic, a customer, a quote or a result.
  Write [NEEDS SOURCE] where a fact is missing.
- Write [N] variants, each labeled with the hook type used:
  claim / number / situation / contrarian / story.

Wait for the platform and the post brief.

Why it works: the labeled hook types do the same job as angle-labeled pairs in my email subject line prompts. You learn which hook type works for your audience rather than which single line did, and that lesson compounds across every post you write afterwards. The “no questions” rule for the first line is deliberate. Rhetorical questions are the model’s default hook, and the fold makes them cheap to scroll past.

The platform prompts

Paste one of these after the base prompt, then the post brief.

1. LinkedIn

Platform: LinkedIn, personal profile of [name, role].
Post brief: [the one thing this post says, and the job: reach /
discussion / leads].
Length: 120 to 200 words. First line under 12 words, the rest in
short paragraphs of one or two sentences with a blank line
between them.
No hashtags in the body. Add up to three at the very end.
End with a specific question only if the job is discussion, and
make it one a reader can answer in a sentence.
Write 3 variants: story hook, contrarian hook, number hook.

Why: LinkedIn rewards a first line that reads like the start of an argument rather than an announcement, and rewards white space, because the post is read on a phone in a feed of dense paragraphs. The three-hashtag cap keeps the post from looking like a broadcast. The question rule exists because “What do you think?” at the end of every post is the most ignored sentence on the platform.

2. Instagram

Platform: Instagram feed post. The image shows: [describe it].
Post brief: [what the caption adds that the image does not, and
the job: saves / shares / profile visits / sales].
First line: under 10 words, no emoji, works as a standalone.
Body: 40 to 120 words. Emoji allowed, max two, never as bullets.
Hashtags: a block of [5 to 10] after the caption, mixing broad
and niche, no banned or generic ones. List them separately so I
can check them.
CTA: one, matched to the job. "Save this" for saves, "send this
to" for shares. No "link in bio" unless the job is profile visits.
Write 3 variants: situation hook, number hook, claim hook.

Why: the image already did the stopping, so the caption’s first line has to add information rather than describe the picture. Matching the CTA to the job is the part most caption prompts skip. “Save this” and “send this to a colleague” ask for different actions, and asking for the one you want is the whole point of writing a caption instead of posting the image alone.

3. X

Platform: X. Single post, no thread.
Post brief: [the one claim, and the job: reach / replies / clicks].
Must be readable in full without expanding. No hashtags. No
emoji. One link at most, at the end, only if the job is clicks.
Write 5 variants under [platform's current limit] characters:
claim, number, contrarian, situation, and one that is a plain
statement of the fact with no hook at all.

Why: X punishes the visible hook harder than any other platform, because the audience has read every hook. The “no hook at all” variant is there because it wins more often than you would expect: a specific fact stated plainly reads as confidence. The no-hashtag rule is not a style choice. Hashtags on X read as a broadcast, and broadcasts get scrolled.

4. TikTok and Reels

Platform: [TikTok / Instagram Reels]. The video shows: [describe
it, and the spoken hook in the first three seconds].
Post brief: [what the caption adds, and the job: watch-through /
follows / comments / search discovery].
First line under 8 words, restating the video's promise in
different words from the spoken hook.
If the job is search discovery, include [2 to 3] phrases people
would type to find this, written as natural sentences, not a
keyword list.
Hashtags: [3 to 5], listed separately.
Write 3 variants: situation, number, contrarian.

Why: short-video captions are read after the video has started, so they should not repeat the spoken hook word for word. The search-discovery instruction is there because both apps now surface videos in search, and a caption that says what the video is about in plain language is the cheapest discovery lever you have. For the script inside the video, the YouTube script prompts run the same promise-keeping logic at long form.

5. Facebook

Platform: Facebook page post. Attached: [link / image / nothing].
Post brief: [the one thing, and the job: clicks / shares /
comments / event responses].
Length: under 80 words unless the job is comments, then up to
150 with a concrete prompt for the comment.
No hashtags. Emoji: none.
If a link is attached, the caption must not repeat the link
preview's headline.
Write 3 variants: situation, claim, story.

Why: the “don’t repeat the preview headline” rule is the one this prompt exists for. The most common Facebook caption is the article headline restated, which tells the reader nothing the preview card did not, and wastes the only line they will read.

The working prompts

The platform prompts write one caption. These are the ones I actually run week to week.

6. One post into a week of captions

Here is a piece of content we published: [paste the post, or the
summary and the three most specific points in it].

Extract five distinct angles from it, each a different claim or
example, not a paraphrase of the same point. For each angle,
name the platform it suits best and why, then write the caption
using that platform's rules above.

Do not reuse a first line, a statistic, or an example across the
five. If the source does not contain five distinct angles, say
so and give me fewer.

Why it works: the “five distinct angles” and “do not reuse” rules are what stop repurposing from becoming repetition. The permission to return fewer matters. A model told to produce five will produce five, and the fourth and fifth will be the first two rephrased. This prompt is the social-media end of the content repurposing stage in a content marketer’s pipeline - and if you want the whole pipeline, extraction pass included, that is the repurposing workflow.

7. The customer quote caption

Here is a real, approved customer quote: [paste, with the
customer's name and company as we are permitted to show them].

Write 3 captions for [platform] that put the quote's most
specific sentence in the first line, in the customer's words,
with attribution immediately after. Do not paraphrase the
quote. Do not add claims the quote does not make. The rest of
the caption gives context in under 40 words.

Why: the instruction not to paraphrase is the guardrail. Models improve quotes, and an improved customer quote is a fabricated one. Putting the customer’s specific sentence above the fold works because the specificity is theirs, and readers can tell.

8. The launch caption set

We are launching [what it is, for whom, and the one problem it
solves] on [date].

Write a three-post sequence for [platform]: teaser (the problem,
no product), launch day (the product, one benefit, one CTA),
follow-up (one concrete example of it in use, [NEEDS SOURCE]
where we have not supplied one).
Each post must be readable without the others. No "as we
announced yesterday".

Why: launch sequences fail when post two assumes the reader saw post one. Almost nobody sees all three, so each has to stand alone. The teaser rule of “the problem, no product” is there because the model’s instinct is to announce in every post, and a teaser that announces is just a launch post with less information.

9. The rewrite-from-results prompt

Here are our last 15 captions on [platform], with the hook type
used and the results: [paste, with reach, saves, comments, or
whatever we measure, and our typical baseline].

Tell me which hook types beat the baseline and which lost, and
which hook types we have not tried. Then rewrite the three
weakest captions using the winning hook type, and write two new
captions using a hook type we have not tried.

Why: this is the prompt that turns the hook labels into a strategy. Models read patterns in performance data well when the data is labeled, and the “hook type we have not tried” instruction keeps the account from settling into one formula that slowly stops working.

Keeping score

The hook labels are only useful if you record them. A four-column log is enough: date, platform, hook type, result against baseline. A month of entries tells you something no caption tip ever will: whether your audience stops for numbers, situations or contrarian claims. That sentence goes into the base prompt as context, and the prompts start out ahead.

Record results as beat baseline or missed baseline rather than raw numbers unless your reach is large. Small accounts produce vivid differences that are noise, and a log full of noise teaches confident nonsense.

The rule above the prompts

Nothing in a caption can promise what the content does not deliver. A hook that oversells earns a tap, a scroll-back and a slow loss of trust in the account name. The base prompt bans invented statistics and quotes for the same reason it bans questions in the first line: both are the cheap way to a first line, and the account pays for the cheap way later.

FAQ

Should I write one caption and adapt it, or one per platform? One per platform, from the same brief. Prompt 6 does this in one pass. An Instagram caption trimmed for X reads like an Instagram caption trimmed for X.

How many hashtags actually help? Fewer than most guides say, and none on X or Facebook in my prompts. On Instagram and TikTok, use them for discovery and keep them out of the sentence. Test the count like any other variable, one change at a time.

Can I use these prompts for a scheduling tool’s built-in AI? The rules transfer. The built-in assistants usually let you paste a system-style instruction, and the base prompt is that. The platform prompts are what most built-in assistants are missing.

What about comments and replies? Different job, different prompt. A reply that reads like a caption reads like a bot. Write replies by hand, or use the voice block alone with “reply in one or two sentences, as a person”.

Also in this library: ChatGPT prompts for blog posts, email subject line prompts and ad copy prompts for Google, Meta and LinkedIn. All of them assume the brand voice system prompt is in place. If captions are one slice of a job you own end to end, the social media manager’s daily playbook shows where these prompts sit in the week. For the wider prompt library and where it fits, start with the complete guide to AI marketing.