Which tool to choose
There's no best tool in absolute terms, there's the right one for how you post. Choose based on how much work you do.
You post now and then, for a single profile (yours or a small brand's) → ChatGPT or Gemini, free version. They write captions, rewrite boring text, and pull out new angles: you give them a well-built prompt and they do the rest, like a junior copywriter always on call. The real limit: on ChatGPT's free plan you have access to the most powerful model only for a certain number of messages before the automatic downgrade, which I discuss further down.
You want text and graphics in the same place → Canva with Magic Write. The writing tool lives inside the editor: you create the image and the caption without jumping from one app to another, with the text already designed to fit in the graphic. Watch the free limits: the Free plan gives limited access to AI tools, about 25 Magic Write prompts a month, while for full, unlimited use you need the Pro plan at 15 dollars a month.
You manage multiple accounts and also have to schedule the posts → Metricool. Here the AI is one piece of a complete flow: you plan, publish, analyze, and generate the monthly report in the same environment. On top of that it suggests optimal publishing times based on your specific account's historical engagement data, not on generic industry benchmarks. The downside: it doesn't produce visual content, for graphics and video you need separate tools.
My operational advice: start with free ChatGPT for the text, and add a second tool only when the volume of work justifies it.
How to do it
The principle is one: the AI writes well only if you give it a precise brief. A vague prompt produces generic captions to throw away. Here's the flow that works, identical on PC and smartphone (from a browser or the app the path is the same).
- Open the chosen tool and start a new conversation.
- Paste the prompt below, replacing only the parts in quotes with your real data.
- Read the three variants and choose the one closest to your tone.
- Ask for a targeted refinement ("shorten the first one", "make the CTA more direct").
- Copy, paste into the post, and change at least one sentence by hand so it sounds like you.
The operational syntax for an Instagram caption:
Act as a social media copywriter.
Write 3 caption variants for an Instagram post.
Post topic: a new line of vanilla-scented candles, handmade
Brand: small artisan workshop, warm and familiar tone, uses informal "you"
Goal: get people to comment with their favorite scent
Audience: women 25-45 interested in home and wellness
Requirements:
- Maximum 4 lines each
- Open with a hook in the first line (no "Here's our...")
- Include a single question as a call to action
- Add at the bottom 8 relevant hashtags, a mix of popular and niche
- Use at most one emoji per caption
Feedback: if the three versions look too much alike or sound like a brochure, the brief was weak. Add a concrete detail (a specific scent, a real customer, a charming flaw in the product) and try again. The more meat you give, the more recognizable the caption becomes.
To adapt the same content to another platform without rewriting from scratch:
Take the caption you just wrote and adapt it to LinkedIn:
more professional tone but not rigid, no emoji,
first line that works as a hook before the "...see more",
close with a question that invites discussion among industry colleagues.
A concrete example
Marta runs the Instagram page of an artisan gelato shop and had to announce a new Bronte-pistachio flavor. By hand it took her twenty minutes and out always came the usual "New flavor available today!".
She opened free ChatGPT and pasted the prompt above, changing the topic to: "new Bronte DOP pistachio flavor, made in limited quantity, available only this weekend". Goal: create urgency and get people into the shop.
In fifteen seconds out came three variants. The first opened with a sentence about pistachio "that costs more than gold" — a strong hook. Marta kept it, shortened the second line, replaced a generic hashtag (#gelato) with two local ones (#gelateriaroma #pistacchiodibronte) and changed the CTA to "Come by before it runs out, or rethink it all week long?". Total time: three minutes. The post got more comments than usual, almost all of them tagging a friend. The difference wasn't made by the AI alone, but by the AI plus the three human touch-ups.
When it does NOT work (and how to fix it)
If the captions all sound the same and fake
This is the number one problem. By default the AI writes in "marketese": enthusiastic, generic, full of empty adjectives. The remedy is to give it raw, real material. Instead of "write a caption about our coffee", write "our coffee comes from a single plantation in Guatemala, we roast it every Tuesday, it has notes of chocolate and orange". Specific details are the only antidote to genericity.
If the hashtags are invented or irrelevant
Language models don't have real-time post counts per hashtag, so they can propose dead or nonexistent ones. Don't trust them blindly: take the 3-4 hashtags that convince you and verify them by hand in the platform's search bar, where you see the real number of posts. For those who work seriously on hashtags, dedicated tools with built-in verification (shadowban check, performance) beat any LLM on this specific point.
If ChatGPT blocks you mid-job
On the free plan it happens. Free users have access to the flagship model as the default, the same as paying users but with lower message limits: when the limit is reached or during high-traffic hours, the system silently switches to the mini model. OpenAI doesn't publish exact numbers: based on community tests as of April 2026 the talk is of about 15-40 messages per 3-hour window, but the limit depends on server load, account age, and message complexity. The zero-cost fix: use the mini model as a fallback, wait for the reset, and in the meantime lean on Microsoft Copilot or Google Gemini. Creating multiple accounts to get around the limit isn't a clever shortcut: it violates OpenAI's terms of service and you risk permanent suspension, including the main account.
If you run out of Magic Write credits on Canva
The free plan runs out fast if you write every day. Practical workaround: write the captions elsewhere (free ChatGPT or Gemini, which on pure text are more generous) and use Canva only for the graphic, pasting the text in by hand in the editor. You save the AI credits for when you really need them.
A tip from someone who actually uses it
Build yourself a "brand prompt" and always reuse it. Just once, write to the AI who you are: tone of voice, words you use and words you'll never use, audience, and two or three of your best already-published captions. Save this block in a note on your phone. From then on, every time you ask for a caption, you paste in that block first and then the topic of the day. The difference between a "generic AI" caption and one "that looks written by you" lies almost entirely here. It's the thing that separates those who use AI well from those who complain that "it always writes the same way."
Frequently asked questions
Which free AI writes the best captions?
For pure text, ChatGPT and Gemini are the most versatile and generous on the free plan. If you want to write and lay out the graphic in the same place, Canva with Magic Write is worth it, keeping in mind that the free plan includes a limited number of AI credits for Magic Write and for image generation.
How long does it really take to make a post with AI?
From one to four minutes for the caption, once you have a good saved prompt. Most of the time goes into the final touch-ups, not the generation. If it takes you longer, it's because the initial brief was too vague and you're chasing the AI with "no, not like that."
Should I tell the AI which platform I'll publish on?
Always. Each platform has different lengths, tones, and hashtag rules, and the AI knows them: stating the destination changes the text's structure. A caption designed for Instagram, pasted onto LinkedIn, jars; and vice versa.
Are AI-written posts penalized by the algorithms?
It's the most widespread fear, but it's misplaced. Algorithms don't "sniff out" AI: they reward content that generates real interaction, namely saves, comments, and shares. A mediocre caption gets penalized whether an AI or a human wrote it. The problem isn't the tool, it's publishing generic text without touching it up. Use the AI for the first draft in 30 seconds, then invest two minutes making it specific and human: it's that refinement that the algorithm, indirectly, rewards.