AI-Powered Autonomous Social Video Content Generation and Optimization System
Clip title: Claude Code + Karpathy’s Autoresearch = GOD MODE!
Author / channel: AI Andy
URL: https://www.youtube.com/watch?v=vjJwgXsMfjM
Summary
The video details the creation of an AI-powered “Content Machine” capable of automatically generating, publishing, and continuously improving social media video content. Inspired by Andrej Karpathy’s “autoresearch” repository, the creator adapted this concept of autonomous experimentation and self-optimization from machine learning code to the domain of viral content generation. The core idea is to create five videos daily, publish them, gather real performance data, identify what worked and what didn’t, and then automatically rewrite the AI’s prompts to generate better content every single day.
The first step involved understanding Karpathy’s original “autoresearch” framework, which takes three inputs: a file to change, instructions on what to optimize, and a method to measure improvement. The creator realized his content pipeline possessed these elements: the script (prompt), the workflow (execution), and social media views (measurable output). He then used an AI assistant to strategize how to integrate real social media views as the evaluation metric. This led to a two-layer system: a “fast loop” for AI-judged criteria (like hook quality) and a “slow loop” for weekly human analysis of actual view data. Through the Meta Graph API, he pulled extensive Instagram and Facebook data, revealing clear patterns: “wow factor” visual demos crushed it, while productivity-focused content underperformed.
A critical component was building robust, objective evaluation criteria. Instead of subjective assessments like “Is this engaging?”, the creator developed 10 binary (yes/no) questions for Gemini, focusing on aspects like hook quality (e.g., “Does the hook describe a RESULT or TRANSFORMATION?”), content framing, emotional/visual impact, and differentiation. These clear, machine-readable questions provided a measurable score out of ten for each script, eliminating ambiguity and enabling precise feedback for the AI.
This formed the “Autoresearch Feedback Loop.” It begins with Data Collection (Meta Graph API pulls latest views), followed by Match & Update (views matched to Airtable records). Then comes Evaluation, which includes Pre-scoring New Ideas (source tweets evaluated by Gemini, high-scoring ones enter a creation queue) and Scoring Published Scripts (Gemini scores actual scripts). Next is Intelligence, where the system Correlates scores with real views to identify winners, false positives, handle rankings, and track approval rates. Finally, Output involves Generate & Push, where Gemini writes improved prompts based on the correlation data and pushes them to the workflow. This self-improving loop runs every 24 hours.
Over two days, the prompt underwent five meaningful revisions, evolving from a simple announcement to one designed to trigger curiosity, hint at dramatic shifts, deepen personalization, and create emotional impact. This continuous evolution is logged, providing a historical record of prompt changes and their effects. The system runs automatically daily, pulling views, updating scores, and generating reports. A human team member provides crucial feedback by reviewing generated videos and marking them as “schedule” or “no,” allowing the AI to learn from human quality inspection. The creator highlights that this system continuously improves, logging every change and its associated cost, thereby creating a valuable “research log” that can be handed to future, smarter AI models. The entire Content Mate system, including the N8N template and autoresearch loop, is offered for free, demonstrating the broad applicability of this self-optimizing framework to any process with a measurable output.