AI Agent for Content Creation: 2 Real Systems

I record a podcast episode. That's it. That's my only job. Everything else (clips, social posts, newsletters, analytics, community management) is handled by AI agents running on a Mac Mini in my office.

And I'm not the only one. Oliver Henry turned an old gaming PC into an AI agent called Larry that generated millions of TikTok views in a single week. Cost per post: about $0.50 in API calls. Time Oliver spends per post: 60 seconds.

AI Content Pipeline: How 13 Agents Turn One Podcast Into Unlimited Content

A single podcast episode generates: 3 video clips per day on X, a newsletter teaser, YouTube metadata optimization, Skool community content, and analytics tracking across all platforms. None of this requires me to touch a keyboard after recording.

Before I built this system, my weekly content workflow looked like this: record the episode, spend 4 hours pulling clips, write 3-5 social posts manually, draft the newsletter, update the YouTube description, and check analytics across 5 platforms. That's roughly 8-10 hours per episode on top of the actual recording. Now? I record. That's it. The agents handle the rest overnight.

Here's the full pipeline, step by step.

How AI Agents Detect and Process New Podcast Episodes

Jimmy (the YouTube agent) checks the podcast RSS feed twice a week at 1 AM. When a new episode appears, Jimmy downloads the full transcript, stores it, and notifies Marc that new material is ready. That one detection triggers the entire downstream pipeline.

Jimmy doesn't just grab the transcript. He also pulls episode metadata: title, description, guest name, timestamps. He stores everything in a structured folder (transcripts/YYYY-MM-DD-guest-name/) so every downstream agent knows exactly where to find the source material. No Slack messages. No handoffs. Just files in the right place.

The detection itself runs as a cron job:

openclaw cron add --schedule "0 1 * * 1,4" \
  --prompt "Check podcast RSS for new episodes. Download transcript. Notify Marc." \
  --agent jimmy

Mondays and Thursdays at 1 AM. Matches our publishing schedule. If no new episode exists, Jimmy logs "no new episode" and goes quiet. Zero wasted tokens.

Using AI to Find the Best Clip-Worthy Moments

Claude (the copy editor) reads the transcript every night at 2 AM and selects the 3 best moments. Criteria: concrete numbers, contrarian takes, tactical insights, emotional stories. Claude writes the X posts and pushes everything to Notion with exact timestamps.

Here's what Claude actually scores on a 1-10 scale for each potential clip:

• Specificity (1-10): Does it contain real numbers, real names, or a concrete process? Vague motivational quotes score 2. "We went from $3K to $47K MRR in 6 months by doing X" scores 9.
• Controversy (1-10): Does it challenge something people believe? Contrarian takes perform 3x better on X than consensus opinions.
• Standalone value (1-10): Can someone watch this 60-second clip and get value without knowing the full episode? If it needs context, it's a bad clip.
• Hook strength (1-10): Does the first sentence grab attention? Claude rewrites the opening line if the natural starting point is weak.

Only moments scoring 7+ across all four criteria make the cut. Before this scoring system, we were posting average clips that got 50-200 views. After? Average view count jumped to 1,500+ per post. The scoring system is the difference.

Automated AI Video Editing with Subtitles

Adrien takes Claude's timestamps and produces broadcast-ready clips: download via yt-dlp, Whisper transcription for word-level subtitle timing, smart cuts to remove filler words, silence removal, and subtitle burn-in (white text, 4-5 words per line max).

The editing pipeline runs in this exact order:

1. Download the segment. yt-dlp grabs only the timestamp range. No downloading the full 60-minute episode for a 45-second clip.
2. Whisper transcription. Word-level timing so subtitles sync perfectly. We use the large-v3 model for accuracy.
3. Filler removal. Adrien detects and cuts "um," "uh," "you know," "like" when they add nothing. This alone makes clips feel 2x more professional.
4. Silence trimming. Gaps longer than 0.8 seconds get shortened to 0.3 seconds. Keeps the pacing tight.
5. Subtitle burn-in. White text, black outline, 4-5 words per line, bottom-center. No fancy animations. Clean and readable on mobile.
6. Final compress. H.264 encoding, optimized for Telegram delivery and X upload. File size under 15MB for quick sharing.

Total processing time: about 90 seconds per clip on the Mac Mini. Adrien delivers the finished clip to a shared folder and pings Bob for QA.

AI Quality Gate Before Publishing

Bob runs QA at 4 AM. Subtitles match what the guest says? Timestamps correct? Writing style followed? Video attached? Nothing ships without passing the gate.

Automated Social Media Scheduling with AI

Dan posts 3x daily to @profitfounder via the Typefully API: 4 PM Bali (EU evening), 9 PM Bali (US morning), 4 AM Bali (US evening). After each post, Dan adds a CTA reply with YouTube + Spotify links.

Why these specific times? Loop (our analytics agent) ran a 30-day analysis of when our audience is most active. The data showed three clear engagement peaks corresponding to European evening scroll time, US morning commute, and US evening wind-down. Posting at these windows gets 2-3x more impressions than random timing.

Dan's posting workflow is fully hands-off:

1. Claude pushes approved clips + copy to Notion each night
2. Dan picks up the next scheduled piece from the queue
3. Dan uploads the video via Typefully's 3-step media upload API
4. Dan schedules the post for the next available slot
5. 60 seconds after the post goes live, Dan adds a reply: "Full episode: [YouTube link] | Listen: [Spotify link]"

The CTA reply trick is important. X's algorithm sees the reply as engagement, which boosts the original post's visibility. And the links don't suppress reach because they're in the reply, not the main tweet.

AI Newsletter Writing from Podcast Transcripts

Tyler drafts the newsletter on episode days. The rule: teasers only. Bullet points of what they'll learn. Never give away the content. Drive clicks to the episode.

Tyler follows a strict template:

• Subject line: Guest name + their most impressive metric. Example: "David Park built Jenni AI to $2M ARR. Here's what he told me."
• Body: 3-4 bullet points teasing the best moments. Each bullet creates curiosity without satisfying it.
• CTA: Single link to the YouTube episode. No multiple links. No distractions.

Tyler pushes the draft to Notion. I review it in 2 minutes and send through Beehiiv. The newsletter open rate sits at 45%+. Tyler never writes more than necessary. Short newsletters that tease great content beat long newsletters that give everything away.

AI-Powered Analytics Across All Platforms

Loop scrapes YouTube, X, Instagram, Skool, and Beehiiv daily. Every Sunday: correlation analysis showing which content drove growth. This data feeds back into Claude's moment selection. The pipeline learns from its own data.

Loop tracks these metrics daily across every platform:

Platform	Metrics Tracked
YouTube	Views, watch time, CTR, subscriber delta, comment count
X (@profitfounder)	Impressions, engagements, profile visits, follower delta
Instagram	Reach, saves, shares, follower delta
Skool	Member count, post engagement, new joins, churn
Beehiiv	Open rate, click rate, subscriber delta, unsubscribes

The weekly correlation report is where things get interesting. Loop identifies patterns like "episodes featuring founders with $1M+ revenue get 40% more YouTube views" or "clips with specific numbers in the first 3 seconds get 2x more X engagement." These insights feed directly into Claude's clip selection criteria the following week.

This feedback loop is what separates a content system from a content machine. Most people publish, check analytics, and then forget. The agents check, learn, and adjust. Automatically. Every week.

Larry: The AI Agent That Got Millions of TikTok Views

My system uses 13 agents across multiple platforms. Oliver Henry proved you don't need any of that complexity. One agent. One platform. Insane results.

How Larry Generates Viral TikTok Content Autonomously

Larry runs on Oliver's old gaming PC (NVIDIA 2070 Super) with Ubuntu and OpenClaw. Larry generates TikTok photo carousels for Oliver's iOS apps. TikTok photo carousels get 2.9x more comments, 1.9x more likes, and 2.6x more shares vs video.

Every slideshow: 6 slides exactly, text overlay on slide 1 with the hook, story-style caption, max 5 hashtags.

The Secret to Consistent AI-Generated Image Slideshows

The challenge: you need the SAME room across all 6 slides in different styles. Oliver's solution: lock the architecture. Larry writes one incredibly detailed room description (dimensions, windows, doors, camera angle, ceiling height) and reuses it across every prompt. Only the style changes between slides.

Larry's Autonomous Content Creation Workflow

1. Larry generates images using gpt-image-1.5 (same model the app uses, so marketing IS the product)
2. Larry adds text overlays with custom code
3. Larry writes a caption with the hook formula: [person] + [conflict] → show AI → mindset shift
4. Larry uploads to TikTok as a draft via Postiz API
5. Larry sends Oliver the caption on WhatsApp
6. Oliver picks a trending sound, pastes caption, publishes (60 seconds)

Larry does 95% of the work. Oliver adds trending audio, the one thing that can't be automated yet.

What Larry Teaches About AI Content Strategy

The biggest lesson from Larry: specificity beats complexity. Oliver didn't build a Swiss Army knife. He built a TikTok carousel machine. One format. One platform. One audience. And he optimized relentlessly for that single use case.

Here's the real insight most people miss: Larry's content IS the product demo. Every slideshow showcases what Oliver's app can do. The marketing and the product are the same thing. That's why the $714/month MRR follows naturally. People see the room designs, want the app, download it.

If you're building an AI content agent, don't try to post everywhere on day one. Pick one platform. Nail one format. Scale from there.

AI Content Creation Principles: What Makes Both Systems Work

1. The human does what only a human can do. I record episodes. Oliver picks trending sounds. Everything else is automated.
2. Quality gates prevent garbage from going live. Bob checks my clips. Oliver reviews drafts in TikTok.
3. Memory makes agents smarter over time. Both systems log lessons and learn from their own data.
4. Skills encode the workflow. My agents have SOPs. Larry has a 500-line skill file. Knowledge lives in files, not repeated prompts.
5. Cron schedules are the real power. Having agents work on a schedule without asking is transformative.
6. Start simple, build up. Neither of us launched the full system on day one.

AI Content Creation Cost Comparison

	My Setup (13 agents)	Larry (1 agent)
Hardware	Mac Mini M4: $700	Old gaming PC: $0
Monthly API	~$100-150/mo	~$15-30/mo
Tools	Typefully + Beehiiv (free)	Postiz (~$20/mo)
Results	3 posts/day + newsletter + analytics	Millions of TikTok views, $714 MRR

Neither of us pays for a social media manager, video editor, newsletter writer, or content strategist.

The Real ROI of AI Content Agents

Let's put real numbers on this. Before AI agents, running my content operation would have required:

• Social media manager: $2,000-4,000/month
• Video editor (clips): $1,500-3,000/month
• Newsletter writer: $1,000-2,000/month
• Analytics person (part-time): $500-1,000/month

That's $5,000-10,000/month in hiring costs. My AI agents cost $150/month total. Even if they only perform at 70% of what a human team could do (and honestly, the consistency is better), the math is overwhelming.

Oliver's numbers tell the same story. A social media manager running his TikTok would cost $1,500-2,500/month minimum. Larry costs $15-30/month. And Larry never takes a day off, never has creative block, and never decides to change strategy mid-week without telling you.

How to Build Your Own AI Content System

You don't need 13 agents on day one. Here's the progression that works:

Week 1: One agent, one task. Install OpenClaw. Connect Telegram. Give your agent one content job. Maybe it writes your daily X post. Maybe it summarizes articles for you. Start small.

Week 2-3: Add a cron schedule. Set up your first automated task. "Every morning at 8 AM, draft 3 post ideas based on trending topics in my niche." Now your agent works without you asking.

Week 4-6: Add a second agent. Once the first agent is reliable, add a second one for a different task. Maybe a copy editor that reviews what the first agent writes. Or an analytics agent that tracks results.

Month 2-3: Build the pipeline. Connect agents so the output of one feeds the input of another. That's when things get powerful. Transcript becomes clips. Clips become posts. Posts become analytics. Analytics inform next week's clips.

The key: each agent has one job. If you try to make one agent do everything, it gets confused and output quality drops. Specialization is the secret. It's the same reason companies have teams, not one employee doing everything.

Start today at installopenclawnow.com. Your first agent can be running in 10 minutes.

Get Both Content Systems as Templates

Inside OpenClaw Lab, members get both systems as copy-paste templates: the full 13-agent podcast pipeline with every SOP and cron schedule, plus guidance on building Larry-style single-agent content machines. 260+ founders building these systems together.

New to OpenClaw? Start at installopenclawnow.com.

Frequently Asked Questions

Can AI agents create content automatically?

Yes, AI agents can automate content creation including blog posts, social media updates, newsletters, and video scripts. They research topics, write drafts, optimize for SEO, and schedule publishing across platforms without manual prompting.

What is the best AI agent for content creation?

OpenClaw is the best AI agent for content creation because it connects to your tools, runs on a schedule, and executes multi-step workflows. Unlike chatbots, it can research, write, format, and publish content end-to-end without you sitting at the keyboard.

How do AI agents differ from AI writing tools?

AI writing tools like Jasper or Copy.ai generate text when you prompt them. AI agents go further by autonomously researching topics, writing content, formatting it, and publishing it on a schedule. Agents act independently while writing tools require constant input.

Can AI agents repurpose content across platforms?

Yes, AI agents excel at content repurposing. They can take a podcast episode and turn it into a blog post, social media threads, newsletter teasers, and video scripts. Each output is optimized for the specific platform's format and audience.

How much content can an AI agent produce per day?

A single AI agent can produce 3 to 10 pieces of content per day depending on complexity. This includes social media posts, email drafts, blog outlines, and newsletter sections. Running multiple specialized agents multiplies this output significantly.