Lenny Rachitsky's newsletter just walked operators through cloning themselves with Gemini Omni in under 15 minutes. Voice, face, knowledge base, response patterns. Fifteen minutes.
The tools are here. The question is whether the clone runs on infrastructure you own — your knowledge graph, your MCP server, your billing — or whether it sits inside a consumer product where the platform sets the rules.
The Three Extraction Economies
LLM training ingestion. Your old work is already in the set. The UK copyright fight made this plain — 88% of creators wanted protection, the government punted.
Hourly IP dumps. Mercor and similar platforms pay experts to feed training pipelines. One-time rate, permanent loss of the compounding asset.
Expert-owned extraction. You own the graph, the agent, the distribution, the billing, the customer relationship.
The Gemini Omni clone tutorial is option two with a friendlier UI. You're clicking the buttons, but you're building inside someone else's walls.
What You Give Up Inside Their Product
Memory. Every conversation is platform data. You see transcripts. They see cross-user patterns.
Distribution. Discoverable on their terms. Algorithm change or first-party competing agent and your asset evaporates.
Billing. Platform takes a cut. Customer card on file with them. You have a permission slip, not an email list.
Knowledge graph. If it lives in their format, you can't move it. Day you leave, you start over.
The KDS + MCP Stack
Four parts, all yours.
Knowledge Graph. Structured, versioned, queryable. Not a folder of PDFs. Frameworks, decisions, edge cases, reasoning patterns organized for agent retrieval. The agent is replaceable. The graph is not.
Agent Layer. Voice, response style, escalation rules, refusal patterns. Trained on your graph.
MCP Server. Model Context Protocol launched from Anthropic in November 2024 and got donated to the Linux Foundation's Agentic AI Foundation in December 2025. Neutral infrastructure now, not a one-vendor bet. Expose your expertise as an MCP server and any agent — Claude, Gemini, ChatGPT, the next thing — calls it on your terms, your rate limits, your auth.
Distribution and Billing. Your list, your Stripe, your domain, your sequences.
With all four, the 15-minute Gemini clone is a front door, not a business model. Ship a Gemini version, ChatGPT version, Claude version, and one on your own site — all calling back to the same graph and MCP server. You own the spine.
The Math
Mid-tier expert. 10,000 subscribers. $300/hour rate. 60 inbound questions a week. 30% close rate on $50 async consults. $46,800/year on the table.
Inside a consumer product with 30% platform cut and no email capture: net ~$25,000. No customer ownership.
On your own KDS with MCP server: net $46,800, build the list, graph compounds. Three years in, that graph powers enterprise licensing, a course, and an embedded partner co-pilot.
Same 15 minutes at the front. Different asset at the back.
This Week
Audit first. The 20 questions you answer most. The 10 frameworks you teach. The 5 case studies you reference. That's the seed.
Pick portable infrastructure. No graph export means it isn't infrastructure. It's a hostage situation with good UX.
Stand up an MCP server. Even a minimal one. Treat it like a domain name.
Then build the clones. Plural. All calling home to your graph.
I'm writing a book about this — On Whose Terms: The New Expert Economy and the Fight for What You Know. Join the launch list.
Keep Building,
— Matt
Matt Cretzman builds Skill Refinery, the KDS + MCP infrastructure layer for expert-owned AI agents. More at mattcretzman.com.
Top comments (0)