Models Are All the Same—Your Context Is Your Company's Moat
Greg Isenberg listed 49 items for building an AI company, and SIP condensed them into 5 missions. The core idea is to build a company brain and let agents do the work. But a commenter warned: agents will make up a decision you never made, and then it becomes company policy.
Greg Isenberg recently posted a long thread listing 49 items for turning a company into an 'AI-pilled' company. The Startup Ideas Podcast condensed it into 5 missions. The core idea in one sentence: models are the same for everyone; your context is the only thing that truly belongs to you.
To translate that: don't expect the model to give you an edge. What's actually valuable is the customer records, decision rationale, and process playbooks you've accumulated. So the first thing isn't deploying a model—it's building a brain.
1. Build a company brain
Greg's original suggestion: start with a folder in Obsidian full of markdown files that capture how you think, write, and make decisions. This is your 'second brain.' Then expand it into a company brain: customers, calls, pricing, playbooks, past decisions, and the reasons behind those decisions. Agents read these files before every task.
A commenter in the replies warned: a second brain is only useful if it is continuously updated. Most teams skip the step of having agents archive to it, and then the folder goes stale after a month.
2. Capture every conversation
Record all calls and meetings (with Granola), and use voice notes throughout the day (Wispr). Every transcript goes into the brain. That way, agents don't start from zero; they work from every real conversation that has happened.
3. Record your best workflow once
Take the workflow that makes the most money or saves the most time, do it once, and record yourself talking through it. The agent turns the recording into a skill. One workflow, one skill. Small and boring beats one big prompt.
A commenter said: those 50 real cases are the real asset. Models come and go; the standards stay.
4. Give agents keys
Connect tools with MCP (Gmail, Slack, Stripe, Notion), then give the agent a browser. That way, the agent can take actions, not just give advice.
A commenter added: have the agent color-code charts by category—it looks and reads much better.
5. Close the loop
Write every completed task back into the brain. Every Friday, the agent rewrites the playbook based on the week's results, and you approve the changes.
Note: one commenter ran into a pitfall—the agent recorded a decision you never made, and then it became company policy. So humans keep 10%: approve, take responsibility, and talk to customers.
Greg says this is the whole game. Models aren't the moat, and data isn't entirely either; your context is. And you'd better accumulate it fast, because others are running the same loop.
You can start Mission 1 this weekend. SIP's version is shorter—5 missions, same core.
The original thread includes the following 5 videos:
发布时间: 2026-11-22 04:35