The short version: Claude writes better captions than ChatGPT in my experience, mainly because it doesn't default to exclamation marks and emoji pile-ups, but it still needs real structure in your prompt or you'll get generic fluff. Give it your actual voice, a real constraint, and a specific job for the caption to do, and it stops sounding like every other brand account on your feed.
Why Claude specifically for captions
I use Claude for most of my writing work now, captions included, and the difference between it and other models is tone control. Ask ChatGPT for a "fun, engaging caption" and you'll get three emojis, a rhetorical question, and the word "journey" somewhere in there. Claude tends to hold back unless you explicitly ask for that energy, which means the starting point is closer to how an actual person talks.
That said, Claude is not magic. It still needs a prompt that gives it something to work with. A thin prompt gets a thin caption, regardless of which AI you're using.
The caption prompt structure that works
I build caption prompts around five pieces of information. Skip any of these and you'll notice the output getting vaguer.
- The platform (LinkedIn caption and Instagram caption should never read the same)
- The specific post content (not "a photo of my product", describe what's in it)
- The one thing you want the reader to do or feel
- Your voice, described in concrete terms, not adjectives like "friendly"
- A length limit, because Claude will ramble if you let it
Here's a prompt that uses all five:
"Write an Instagram caption for a photo of our bakery's new sourdough loaf fresh out of the oven, steam still rising. Goal: get people to comment which topping they'd want. Voice: direct, slightly cheeky, no exclamation marks, like you're texting a mate not announcing a product launch. Max 40 words. Give me 3 versions."
A worked example
Say you run a small plant shop and you've just got a delivery of monstera cuttings in. You could ask Claude for "a caption about new plants" and get something like "New arrivals alert! Who's ready to add some green to their space? Shop now before they're gone!" That's the generic version, and it's the one most people settle for because it's fast.
Instead, try this prompt: "Write a caption for a photo of monstera cuttings sitting in water jars on our counter. We only got 14 of them. Voice: dry, a bit deadpan, we don't do urgency language or countdown pressure. Goal: get people to ask if we have more coming. One short caption, under 25 words."
What comes back is closer to: "14 monstera cuttings arrived this morning. That's it, that's the whole batch. Ask us about the waitlist." No exclamation marks, no "shop now", and it reads like a person wrote it standing behind the counter, because the prompt gave Claude a real detail (14, not "some") and a real restriction (no urgency language).
The uncomfortable bit nobody tells you
Most posts on this topic will give you ten prompts and tell you to pick your favourite. What they won't tell you is that if you and five other small business owners all paste in the same generic prompt template, you'll get captions that sound suspiciously alike. I've seen this happen across LinkedIn, where three different consultants post AI-flavoured captions in the same week that all open with a short punchy line followed by a colon. Readers notice patterns faster than you think, even if they can't name what's bothering them.
The fix isn't to avoid AI, it's to feed it things only you know: your actual numbers, your actual customer complaints, the specific phrase your regulars use. Generic prompts produce generic output no matter how good the model is. This is the same principle behind a real structured prompt framework, the quality of what you put in sets the ceiling for what comes out.
Prompts for different caption jobs
Not every caption needs to sell something. Here's how I'd split the prompts by job:
- Announcement caption: "Write a caption announcing [specific thing]. State the fact plainly in the first line, no hype words. Then one line on why it matters to the reader, not to us."
- Engagement caption: "Write a caption that asks one specific question related to [topic], not a generic 'what do you think?'. The question should only make sense if someone has used/seen/tried [product/topic]."
- Behind-the-scenes caption: "Write a caption for a photo of [specific scene]. Include one honest detail that isn't flattering, like a mistake or a messy bit of the process. Voice: self-aware, not self-deprecating."
- Carousel/thread caption: "Write an opening line for a carousel about [topic] that would make someone stop scrolling, without using a question or a number as the hook."
That last one matters more than it sounds. Claude's default hook style, like most AI, leans on numbers ("5 ways to...") and questions ("Ever wondered...?"). Both work, but overuse them and your feed starts to look like a listicle factory. Forcing the prompt to exclude both pushes Claude toward something more specific.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
A step-by-step for batching a week of captions
If you're planning content in advance, which I'd recommend over writing captions the morning you post, here's the process I'd use:
- List your 5 to 7 posts for the week with one line each on the actual content (photo, video, link)
- Write a single "voice brief" paragraph once, covering tone, words you avoid, and sentence length preference, and reuse it in every prompt that week
- Prompt Claude one post at a time rather than asking for all seven in one go, the quality drops when you batch too much in a single request
- Ask for 2 versions per caption, not 5, because past two options you start picking the safest one rather than the best one
- Edit every single caption by hand before it goes out, even if it's just swapping one word for something only you'd say
That last step is the one people skip, and it's the one that matters most. An unedited AI caption is recognisable within a sentence or two. A thirty-second edit pass fixes that almost every time.
Where this fits into your wider content process
Captions are the smallest part of a content plan, and it shows when people treat them as the whole job. If you're building out a quarter of content rather than scrambling week to week, it's worth starting with real content planning prompts so your captions have actual campaigns and themes to sit inside, rather than being written in isolation for whatever photo you happened to take that day.
And if captions are just one piece of a bigger social media workload you're trying to hand off to AI, it's worth looking at the broader toolset too. I've written about the AI tools that can help manage social media beyond just caption writing, and if video is part of your mix, the AI video generator comparisons are worth a read before you commit to one platform.
If you've tried all this and still can't get your captions to sound like you rather than like AI, that's usually a sign the prompt isn't the problem, your process is. That's exactly the kind of thing an AI implementation coach can sort out in an afternoon, by looking at your actual workflow rather than handing you another list of prompts.
Frequently asked questions
Is Claude better than ChatGPT for social media captions?
In my experience Claude produces less generic, less emoji-heavy captions by default, but both models need a detailed prompt with your actual voice and specifics to avoid sounding like a template. The model matters less than the prompt quality.
How long should a Claude caption prompt be?
Longer than you'd think, aim for 3 to 5 sentences covering the platform, the specific content, the goal, your voice described in concrete terms, and a word limit. A one-line prompt like "write me a caption" produces the most generic result possible.
Can Claude write captions in my brand voice automatically?
Not automatically, you need to describe your voice with specifics like sentence length, words you avoid, and punctuation habits rather than adjectives like "friendly" or "bold", which every brand claims and which mean nothing to an AI model.
Should I post AI-written captions without editing them?
No. Even a strong Claude output benefits from a thirty-second edit pass to swap in a word or detail only you would use, since unedited AI captions are often recognisable to readers within a sentence or two.
Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works
