50 AI Agents for Social Media: Beautiful Diagram, Missing Bill

jaffa drew a social media marketing team as a folder tree: 10 departments, each with 5 agents, all under Claude Code. The diagram is very clean. So clean that you might think the hard part is clicking a few times to create 50 folders.

According to jaffa, this is a system he built in three months that generated $99,705. He is also a co-founder of Virlo, and this tutorial is also promoting Virlo. Keep those two things in mind while reading.

The system itself is not mysterious. An agent is a short file in .claude/agents/ that specifies a name, one responsibility, and the tools it is allowed to use. A skill is a folder in .claude/skills/ that holds a department's workflows, templates, task recipes, and checklists. Fifty agents do not equal fifty models; most are the same model with different instructions and tool permissions. What really differentiates them is the input.

Claude cannot scroll TikTok. The original solution is to connect Virlo first. Virlo is a social media intelligence platform that, according to the original post, has data on more than 20 million viral videos across TikTok, YouTube Shorts, and Instagram. It watches videos, compares data, tracks keywords, and then feeds the results to Claude through a connector or Claude Code's MCP. Only then do the agents know which videos are rising this week, which hooks are being reused, and which creators are producing repeated hits. Without this layer of data, 50 agents are just 50 folders that can write copy.

The 30-day build sequence follows the same logic. On Day 1, connect Virlo and write a niche context file. All agents read this page first so they know the category, audience, product, and competitors. On Day 2, use one prompt to create the 10 departments. In Week 1, add research and competitor tracking; in Week 2, work on outlier videos, outlier creators, hooks, and format trends; in Week 3, write scripts and amplify a single winner; in Week 4, produce weekly reports and content operations. The original post gives a coffee brand example: a local cafe's 16-second video about finding a leaking milk pitcher got more than 17 million views, and the team broke down its structure and applied it to oat milk, a jammed grinder, and an order no one would admit to. One winner was turned into a week of content. This example is more useful than the number 50 agents.

The cost section deserves separate attention. Someone in the comments asked about the token bill for each agent and each piece of content; jaffa replied that it is a Claude subscription plus about $10 per month for the Virlo API, and said the system has already been fully shared. That number only covers the tool costs he mentioned; it does not include shooting, labor, trial and error, or content production. The diagram does not draw the bill either. For ordinary users, do not build 50 first. Write one page of context file, connect one data source, run one outlier-video analysis, and see whether the result helps you avoid shooting one bad video. If it does, then talk about departments.

Original link: jaffa's original post. The full article is here.