OpenClaw Larry Skill: Autonomous TikTok Marketing for App Revenue

Treat the AI as a single employee with one job rather than a multi-agent mission-control system.

Oliver runs everything through one WhatsApp thread with Larry; when a long-running task is needed, Larry spins up a sub-agent and passes full context, keeping the main agent free for new requests.

Why: Simplifies mental model, keeps all context in one place, and aligns with OpenClaw’s intended usage.

"I just have Larry as the one agent that I text through WhatsApp. We just message like you would an employee."

Views and conversions are governed by two independent levers: hook quality drives reach, CTA quality drives conversion.

Low views → bad hook; low conversions → bad CTA; high views but low conversions → hook works, CTA fails; low views but high conversion rate → hook fails, CTA works.

Why: Separating these variables lets the agent iterate each component without guessing.

"if we're getting low views, the hook is bad. If we're getting low views and low conversions to the app, the CTA is bad"

Feed real-time analytics back into the agent so it can self-correct faster than any human.

Larry continuously reads TikTok analytics and app-download metrics, then rewrites hooks, images, and even onboarding flows based on what actually drives revenue.

Why: Creates a closed feedback loop that compounds learning and performance without human bottlenecks.

"the Larry loop basically of having your TikTok analytics, the content creation, feed the analytics back into the content creation until you get a winner"

Post as draft from phone, add trending sound manually, then publish—avoids API penalties and boosts algorithm reach.

TikTok throttles API-posted content; posting from mobile signals human authorship and allows adding viral sounds that the API can’t attach.

Why: Algorithmic favor plus sound boosts outweigh the extra 30-second human step.

"TikTok knows if it's posted through an API… If you post it from your phone, it assumes a human's posting it."

Let the agent run creative experiments—even ones that look wrong—because aggregate data beats human taste.

Oliver almost killed a post because text placement looked off; it became his best performer at 400 k views. He then fully delegated creative choices.

Why: Agent has access to thousands of micro-signals humans can’t see; micro-management reduces performance.

"I flamed at him… I wake up, it's on hundreds of thousands of views… this is just helping even more to drive conversions."

Package every capability as a downloadable skill so the agent can be restored or moved anywhere.

Each project has a memory.md file; skills are plain text with full context, enabling backup, versioning, and local hosting without cloud lock-in.

Why: Future-proofs work against model changes, hardware failure, or service shutdowns.

"skills are infinitely powerful, because they're not just a black box… you own that thing… If you don't like something… just ask your agent to change it."

Lock down variables (room size, window placement) before generating images to avoid hallucinations.

Early manual prompting revealed ChatGPT kept adding doors/windows; Oliver froze these parameters in a prompt template that became the app’s core feature.

Why: Reduces user frustration and increases perceived app quality, turning a bug into a product moat.

"I created a lockdown prompt in ChatGBT to keep the room size the same, the windows in the same places… turned it into an app."

Use high-performing TikTok windows to rewrite onboarding and paywalls inside the app.

After Larry saw churn, he rewrote the onboarding flow; new-user spikes appeared within hours, proving traffic-to-product alignment matters more than more traffic.

Why: Maximizes lifetime value from each viral spike instead of just chasing more views.

"he has completely rewritten my onboarding… This got published two days ago, and this is the most new users I've had in a day for a long, long time."