In their own words
See what founders
have to say
How we think about AI-native work, what we run on, what we ship, how we plan, and how it lasts past anyone's departure. Every protocol has its own page.
Content management systems were built for a world where humans typed into boxes. AI generates, structures, and updates content directly. The CMS is now the model plus a folder of context.
Four tools, one operating model. Claude is the AI partner, GitHub is version control, Vercel is deployment, Supabase is the database. Master these four and you can build almost anything.
An agent is not a magical entity. It is a folder: a role, skills, context, routines, guardrails, and a Definition of Done. Structure the folder well and the agent works. Structure it badly and it does not.
A team is always humans plus AI. Your build team is you, an engineer when you are stuck, and four agents that help you build: PM, Developer, QA, DevOps.
Commit, push, pull, branch, merge. Checking in and checking out with a team, a branch per change, and Claude worktrees so several builds run at once without losing each other's work.
How deployments work, how custom domains attach, and the DNS basics you need so a site goes live cleanly the first time.
Tables, rows, columns, relationships, and row-level security. Enough of the data model to read, write, and reason about a database without breaking it.
Always be cataloguing. Folders laid out the same way every time, files whose names say what they are, and what the chat decided written down before the session ends. The boring parts that make everything else work.
Prompt engineering was typing better sentences into a box. It is over. The unit of work is now a plan: an interview, a product plan in your words, a build plan for the AI, and a Definition of Done you can tick. A vague plan gets you average AI slop. A precise plan gets you a system.
What separates an incredibly smart plan from average AI slop is what you put inside it: business and academic frameworks. You do not memorize them. You know they exist and point at them.
Plan well, build small, build all, review. That is the play. Multi-delegating is running it several times at once: one plan per build, one worktree per build, merged in one at a time.
Give Claude a design system and the output stays on-brand and consistent. One page of tokens and components, a voice page, and examples: what to feed in so what comes out matches.
How to send messages programmatically. Resend for transactional email, the shape of an automated nurture sequence, and the deliverability and opt-out rules that are not optional.
Building internal tools (a CRM, an admin UI, the Monday dashboard) so the team can read and act on the data without escalating to engineers. Synthetic data so you can see it full before the real data arrives.
Everything so far ran on your laptop with you in the loop. A cloud agent runs without you: a function on Vercel wakes on a schedule, reads your database, thinks with Claude, acts within its guardrails, and logs what it did.
Working means it does the right thing today. Hardened means it keeps doing the right thing after the next change. The development protocol, the QA plan, unit tests versus the full suite versus a human pass, and a senior code review before anything emails a customer.
End the day by reflecting with the PM agent: what shipped, what failed QA, what slowed you down. It writes the notes and stages tomorrow, so the next morning starts with a plan instead of a hunt for where you left off.
How to hand off agent context, memory, and project state to the next person, so the work survives a transition without a knowledge tax.
At the Infinite Leverage retreat you walk all 18 protocols in 48 hours in the room, building your own site, your own CRM, and eight working agents.
In their own words