How to run
social media
with AI

After all the hype about AI agents, content factories and one-person marketing departments, it is tempting to think that you can open Claude Code or Codex, connect a few tools and let your social accounts carry on by themselves. In the ideal version, you occasionally nod in approval. In reality, Claude can remove a lot of repetitive work — once you explain where the data lives, how your brand sounds and what a good result looks like.

Running social media is not just writing a post. You need to find the source material, adapt it for several platforms, preserve the brand voice, assemble visuals, schedule publication, review analytics and decide what happens next.

The good news is that much of this cycle can already be built around one AI assistant. The less exciting news: there is still no “turn on the marketing department” button. There is, however, a workable stack of a few components.

1. Connect your platforms and analytics

Until Claude knows what you have published and how it performed, it behaves like a very eager intern on day one: plenty of ideas, no context. Start with a service that can give the AI access to social data and, if you allow it, put approved posts into a schedule.

One option is Metricool. It brings metrics from several platforms into one place and supports scheduling. Claude can then turn that data into a short report: what went live, what grew and which strange number deserves a human look.

If Metricool does not fit, alternatives include Composio and other integration services. Ignore the wall of logos and ask three questions: does it support your platforms, does it expose the metrics you need, and can it keep publications as drafts? If there is a free plan or trial, use it in the spirit of “I’m only looking”. Prove it solves the problem before paying for it.

Drafts first.We can enable automatic publishing after the AI stops congratulating the audience on New Year in July.

2. Add a radar for content and competitors

Your own analytics answers “what worked for us”. A market radar answers “what is happening around us right now”. This is not an invitation to copy somebody else’s Reel and replace “therapist” with “marketer”. It is a way to spot growing topics, formats and mechanics before they reach the next trends newsletter.

For example, Sandcastles tracks Reels and TikTok: what competitors published, which videos grew unusually fast and which visual or written hooks may have helped. For YouTube, there is vidIQ. Do not collect every service at once. If your own Metricool analytics is enough, start there.

3. Teach the AI to write for you

A general-purpose model can produce perfectly average copy. Perfectly average copy offends nobody, proves nothing and quietly leaves the reader’s memory around paragraph three.

This is where skills come in: separate instruction sets with rules, examples and checks. One can turn a long recording into posts, Stories, carousels and scripts. Another protects your tone of voice. A third checks a short-video script against proven patterns. A fourth removes wooden phrases such as “in today’s rapidly changing world”.

1. A format skill. Reels, for example, need their own rules for structure, hook, pace, length and closing transition. You can use an existing skill as a starting point and replace its examples with your own. The open Vyral Content Skills repository ↗︎ includes skills for hooks, scripts, short videos and different platforms.
2. A voice skill. It should know your preferred constructions, acceptable irony and sarcasm, banned phrases, paragraph length and the way you explain difficult things. A serious version also needs examples, checks and explicit boundaries.
3. A humanizer. This is a useful final filter for common signs of machine-written prose. It cannot create a personality from nothing: if the original thought is empty, the result will merely be less wooden.
4. A self-check. A good skill compares its output with the rules: did the meaning survive, is the copy specific, are the facts real, and does the format suit the platform?

4. Connect the visual layer

Canva is the most straightforward starting point for covers, cards, carousels and basic layouts. If its connector is available in your version of Claude or Codex, find it under Settings → Connectors and connect your account. The other route is a dedicated design agent trained on your visual identity.

Neither Canva nor an agent can infer a complete brand system from a 240-pixel logo. They need templates, colours, fonts, composition rules and examples of work you consider good. Otherwise the result will be perfectly tidy — like an insurance-company presentation from 2017.

Begin with three or four approved templates. Let AI change the copy and image inside a clear system instead of rediscovering its inner art director every morning. Once that works, read “How to keep an AI-generated presentation on brand” and build your own visual skill.

5. Add marketing skills selectively

Ready-made libraries can provide methods for content strategy, research, SEO, email, campaigns and competitor analysis. The open Rampstack repository on GitHub, for example, contains more than one hundred modular skills for brand, content, growth, research and operations.

The important word is modular. Installing 103 skills will not spontaneously assemble a marketing director. Choose two or three for the next process — perhaps content repurposing, brand-voice review and a weekly report. Test them on your own material, then add the next one.

Where to begin when this feels like a lot

Do not build an autonomous newsroom that monitors eight platforms, generates video and replies to followers in its sleep. Start with one source and one weekly cycle.

Your first task might be: “Take this webinar recording and turn it into one post, three Stories and a short-video script. Preserve the facts and my voice. Prepare a carousel draft using the approved template. Publish nothing without confirmation.” That alone is enough to stop moving the same text between six windows.

What not to hand over without supervision

1. Strategy. AI can suggest options. A person decides what reputation and position the brand is building.
2. Unverified facts. Every number, case and quotation needs a source.
3. Publishing in the brand’s name at the start. Begin with drafts, an error log and human approval.
4. Unnecessary permissions. An agent that only reads analytics does not need the ability to delete posts or change ad campaigns.
5. Sensitive replies. Complaints, crises, personal data and legally significant messages are no place for cheerful improvisation.

So, does this replace a social media team?

For an independent expert, consultant or small company, this setup can replace a significant amount of mechanical work previously spread across several people. It helps repurpose material faster, maintain consistency and stop assembling reports by hand.

Strategy, a strong original idea, audience judgement and final editing do not disappear. Claude can help one person work like a compact team. It cannot turn the absence of a position into content marketing — although the presentation about it will probably look excellent.