What Is an AI Social Media Agent (and Why Should You Care)?
An AI social media agent is software that autonomously creates, schedules, and publishes content to platforms like TikTok, Instagram, LinkedIn, and X without constant human input. Unlike a chatbot you ask questions to, an agent has its own persistent identity, memory, and access to real tools. It can generate images, write copy, analyze performance data, and post content on a schedule.
The market for AI-driven content automation is growing fast. TikTok’s own data shows that photo carousels (slideshows) get 2.9x more comments, 1.9x more likes, and 2.6x more shares compared to standard video content. The algorithm is actively pushing photo content in 2026, making automated slideshow creation one of the highest-leverage strategies available right now.
Oliver’s agent, Larry, proves the concept works at scale. Within one week, Larry generated over $588 in monthly recurring revenue across two iOS apps, simply by automating TikTok marketing that Oliver had been doing manually for years.
How Larry the AI Agent Works: The Full Technical Stack
Larry runs on an old gaming PC with an NVIDIA 2070 Super GPU, wiped and reinstalled with Ubuntu. The hardware requirements are minimal: 2 GB RAM (4 GB recommended), 1-2 vCPUs, and 20 GB SSD storage. A Raspberry Pi or cheap VPS works fine.
The stack has four key components:
- OpenClaw — An open-source AI agent framework that gives Claude (Anthropic’s AI) a persistent identity, memory, and tool access. This is the brain.
- OpenAI’s gpt-image-1.5 API — Generates photorealistic room transformation images at $0.50 per slideshow ($0.25 with Batch API).
- Skill files — Markdown documents (500+ lines) that teach Larry specific workflows, rules, and lessons learned from every failure.
- Postiz — The social media scheduling platform with an API that Larry uses to upload slideshows as TikTok drafts automatically.
The workflow is straightforward. Larry generates six slides using locked prompt templates, adds text overlays with hook copy, writes a story-style caption, and uploads everything to TikTok via the Postiz API. Oliver’s only job is opening TikTok, adding a trending sound, and hitting publish — about 60 seconds of work per post.
Why Postiz Was Chosen for Automated TikTok Posting
Oliver chose Postiz for a specific reason: it has an API that’s included in the price, with documentation simple enough for an AI agent to understand and use without human intervention. When you’re building autonomous systems, the quality of your API documentation matters as much as the API itself.
Postiz supports 28+ social media platforms from a single dashboard, including TikTok, Instagram, LinkedIn, X, YouTube, Reddit, Pinterest, Threads, and more. For Larry’s use case, the TikTok Content Posting API was the critical feature — it allows uploading slideshow images as draft posts programmatically.
How the Postiz API Handles TikTok Slideshows
Here’s how an AI agent posts a TikTok slideshow through the Postiz API. First, you upload each slide image:
postiz upload ./slide1.jpg
# Returns: {"id": "abc123", "path": "https://uploads.postiz.com/slide1.jpg"}
Then you create the post with all slides attached:
curl -X POST https://api.postiz.com/public/v1/posts \
-H "Authorization: your-api-key" \
-H "Content-Type: application/json" \
-d '{
"type": "schedule",
"date": "2026-04-17T14:00:00.000Z",
"posts": [{
"integration": {"id": "your-tiktok-id"},
"value": [{
"content": "My landlord said I can't change anything...",
"image": [
{"id": "slide1", "path": "https://uploads.postiz.com/slide1.jpg"},
{"id": "slide2", "path": "https://uploads.postiz.com/slide2.jpg"},
{"id": "slide3", "path": "https://uploads.postiz.com/slide3.jpg"},
{"id": "slide4", "path": "https://uploads.postiz.com/slide4.jpg"},
{"id": "slide5", "path": "https://uploads.postiz.com/slide5.jpg"},
{"id": "slide6", "path": "https://uploads.postiz.com/slide6.jpg"}
]
}],
"settings": {
"__type": "tiktok",
"privacy_level": "SELF_ONLY",
"duet": true,
"stitch": true,
"comment": true,
"content_posting_method": "DIRECT_POST"
}
}]
}'
Setting privacy_level to SELF_ONLY means the slideshow lands in your TikTok drafts folder. This is intentional — music is everything on TikTok, and you can’t add trending sounds via any API. The agent handles 95% of the work; the human adds the finishing touch that can’t be automated yet.
Using the Postiz MCP Server for AI Agents
For AI agents built on frameworks like Claude Desktop or OpenClaw, Postiz also offers an MCP (Model Context Protocol) server. This gives your agent direct access to eight specialized tools:
- integrationList — Discover all connected social accounts
- integrationSchema — Get platform-specific posting rules and character limits
- schedulePostTool — Schedule, draft, or immediately publish posts
- generateImageTool — Generate AI images directly within Postiz
- generateVideoTool — Create videos for post attachments
- triggerTool — Execute platform helpers (get Reddit flairs, YouTube playlists, etc.)
Setting up the MCP connection is a single configuration entry:
{
"mcpServers": {
"postiz": {
"url": "https://api.postiz.com/mcp/your-api-key"
}
}
}
With this connected, your AI agent can discover available platforms, understand each platform’s posting rules, generate images, and schedule posts — all without writing custom API integration code.
The Prompt Engineering That Made It Work
The hardest technical challenge wasn’t the API integration. It was getting AI-generated images to look consistent across six slides in a slideshow.
Oliver’s apps — Snugly (AI room makeovers) and Liply (lip filler previews) — required photorealistic output. Early attempts with vague prompts like “a nice modern kitchen” produced completely different rooms on every slide. Windows appeared and disappeared. Counters moved walls. It looked fake because it was fake.
The solution: lock the room architecture and only change the style.
Larry writes one obsessively detailed room description — dimensions, window count, door location, camera angle, furniture size, ceiling height, floor type — and copies it into every single prompt. Only the style changes between slides: wall color, bedding, decor, lighting fixtures.
Here’s a real prompt example:
iPhone photo of a small UK rental kitchen. Narrow galley style kitchen, roughly 2.5m x 4m. Shot from the doorway at the near end, looking straight down the length. Countertops along the right wall with base cabinets and wall cabinets above. Small window on the far wall, centered, single pane, white UPVC frame, about 80cm wide. Natural phone camera quality, realistic lighting. Portrait orientation. Beautiful modern country style. Sage green painted shaker cabinets with brass cup handles. Solid oak butcher block countertop…
The bold section is the only variable. Everything else stays identical across all six slides. This technique produces transformations that look like the same room redesigned in different styles — exactly what users see when they download the app.
The Viral Hook Formula: What Actually Gets Views on TikTok
Larry’s first hooks bombed. “Why does my flat look like a student loan” got 905 views. “See your room in 12+ styles before you commit” got 879. “The difference between $500 and $5000 taste” managed 2,671. All dead.
The problem? Every hook was self-focused. Talking about features, problems, or the app itself.
Then they tried: “My landlord said I can’t change anything so I showed her what AI thinks it could look like.” That post hit 234,000 views — more than everything else combined.
The formula that emerged:
[Another person] + [conflict or doubt] → showed them AI → they changed their mind
Every post following this pattern clears 50,000 views minimum. Most clear 100,000. The hook works because it creates a tiny story in the viewer’s head before they even swipe. You picture the landlord’s face. You picture the reaction. It’s not about the app — it’s about the human moment.
Real examples that worked:
- “My landlord said I can’t change anything so I showed her what AI thinks it could look like” — 234,000 views
- “I showed my mum what AI thinks our living room could be” — 167,000 views
- “My landlord wouldn’t let me decorate until I showed her these” — 147,000 views
How the Agent Learns and Improves Over Time
Most AI automation setups fail because they’re static. Larry isn’t. He has two types of persistent knowledge:
Skill files are markdown documents that teach specific workflows. Larry’s TikTok skill file is over 500 lines and has been rewritten roughly 20 times in the first week alone. Every rule about image sizes (1024×1536 portrait, always), text overlay positioning (6.5% font size, not 5%), hook formats, caption formulas, and hashtag strategy lives here.
Memory files log every post, every view count, and every insight. When Larry brainstorms hooks, he references actual performance data — not guesses. This compound learning is the real advantage. Every failure becomes a rule. Every success becomes a formula.
Planning happens in batches. Oliver sits down with Larry and brainstorms 10-15 hooks at once, reviewing performance data and picking the best ones for the next few days. Larry pre-generates everything overnight using OpenAI’s Batch API (50% cheaper than real-time), so an entire day’s content is ready by morning.
When you’re building systems that manage recurring tasks like social media posting, having good scheduling and automation tools matters.
The Failures That Made It Work
Oliver and Larry tried several approaches before finding what works. Each failure taught a specific lesson:
Local image generation with Stable Diffusion failed. The NVIDIA 2070 Super GPU could run Stable Diffusion, but the image quality wasn’t photorealistic enough for room transformations. The gap between local generation and gpt-image-1.5 was too large, and the API costs ($0.50/post) turned out to be negligible.
Wrong image dimensions killed engagement. Early posts used 1536×1024 (landscape) instead of 1024×1536 (portrait), causing black bars on every video. TikTok is a portrait-first platform.
Unreadable text overlays. Font sizes too small (5% instead of 6.5%), positioned behind TikTok’s status bar, and horizontally compressed because line lengths exceeded the canvas max width.
People in generated images didn’t work. They quickly discovered that AI-generated people in room transformation images created an uncanny valley effect that viewers scrolled past.
Each of these failures was logged in Larry’s skill files as a permanent rule. The agent literally cannot make the same mistake twice because every lesson is encoded in its operating instructions.
Setting Up Your Own AI Social Media Agent: Step by Step
Here’s the exact setup process, with the technical details Oliver used:
Step 1: Get Your Hardware Ready
Any spare computer works. A Raspberry Pi, a cheap VPS, or an old desktop. Install Ubuntu if you’re not sure what to pick. Minimum specs: 2 GB RAM, 1-2 vCPU, 20 GB SSD.
Step 2: Install Your AI Agent Framework
Oliver used OpenClaw, an open-source framework that gives an AI model persistent identity, memory, and tool access. Install it and configure it with your preferred model (Oliver chose Claude by Anthropic).
Step 3: Set Up Image Generation
Sign up for an OpenAI API key at platform.openai.com. You’ll use gpt-image-1.5 for generation. Budget about $0.50 per slideshow, or $0.25 with the Batch API for overnight generation.
Step 4: Connect Postiz for Automated Posting
Sign up at Postiz and connect your TikTok account. Then install the Postiz CLI:
npm install -g postiz
postiz auth:login
Or set up via API key:
export POSTIZ_API_KEY=your_api_key_here
Test the connection by listing your integrations:
postiz integrations:list
# Returns: [{id: "tiktok-abc", name: "Your TikTok", platform: "tiktok"}]
For AI agent frameworks, connect via MCP for the smoothest experience — your agent gets structured access to all posting, scheduling, and analytics tools without writing custom HTTP requests.
Step 5: Write Your Skill Files
This is the most important step. Create markdown documents that teach your agent:
- Image sizes and formats (1024×1536 portrait for TikTok)
- Prompt templates with locked architectural descriptions
- Text overlay rules (font size, positioning, max line length)
- Caption formulas and hashtag strategy (max 5 hashtags on TikTok)
- Hook formats proven to work in your niche
- A failure log so the agent never repeats mistakes
Write them like you’re training a new team member who’s highly capable but has zero context. Be obsessively specific. Include examples. Document every mistake.
Step 6: Start Posting and Iterating
Your first posts will probably underperform. That’s expected. Log what went wrong, update the skill files, and keep going. The system compounds — Larry didn’t start good, but after a week of documented failures and rule updates, he became better at creating viral slideshows than Oliver himself.
The Results: Real Numbers from One Week
Here are the actual metrics from Larry’s first week of autonomous posting:
- 500,000+ total TikTok views in under one week
- 234,000 views on the top-performing post
- 4 posts over 100,000 views each
- 108 paying subscribers across Snugly and Liply
- $588/month MRR and growing
- $0.50 cost per post in API calls ($0.25 with Batch API)
- 60 seconds of human time per post (adding music and publishing)
These aren’t vanity metrics. The views convert into real app downloads, real trials, and real paying subscribers. People watch the slideshow, download the app, try it, and subscribe.
Key Takeaways for Building Your Own AI Content Agent
Oliver’s experiment with Larry reveals several principles that apply to anyone building AI-powered social media automation:
The agent is only as good as its memory. Skill files and failure logs are the difference between an AI that repeats mistakes and one that compounds intelligence. Without persistent, structured knowledge, every session starts from zero.
Hooks beat features. Nobody cares about your product’s capabilities in a TikTok caption. They care about human stories, conflicts, and reactions. The formula “[Person] + [doubt] → showed them AI → mind changed” consistently outperforms feature-focused messaging by 10-50x.
Automate 95%, humanize 5%. Larry handles image generation, text overlays, captions, and posting. Oliver handles music selection and final publishing. This split maximizes both AI efficiency and human creative judgment where it matters most.
Choose tools with AI-friendly APIs. The reason Postiz worked where other tools didn’t is that its API documentation was clear enough for an AI agent to understand autonomously. When building automation, your tools need to be readable by machines, not just humans.
Start bad, iterate fast. Larry’s first posts were “honestly embarrassing” according to Oliver. Wrong image sizes, unreadable text, hooks nobody clicked. But every failure became a rule, and within a week, the system was outperforming years of manual work.
The real unlock isn’t the AI itself. It’s the system you build around it — the skill files, the memory, the feedback loops, and the API integrations that let it operate autonomously. With tools like Postiz handling the multi-platform posting layer and frameworks like OpenClaw providing the agent intelligence, building your own Larry is more accessible than ever.
