A race for Infinite Leverage alumni. One deep dive a week, one build a week, each one adding a layer to your own fork of 8 Edges until your company runs on it. Scored, timed, with a start line and a ship day.
The retreat took you to stage 4: the stack installed, your first app live. The challenge takes you to stage 7, an AI system your team runs. You never start a week from nothing. You fork 8 Edges, a working company OS, load your business into it, and extend it one component at a time.
By week 8 your own 8 Edges is live on your domain, running your business on one central database, with all four offices reading and writing it.
Every prospect and customer in one record. The system knows what each person asked about, bought or ignored, and writes to them accordingly.
People have a journey from first interview to last day. AI team members are planned as folders and built in the cloud, with a role, a routine and a boss.
Workboards, meetings and the Monday brief come from the data, not from someone compiling a report.
Ideas come in from your team through the portal, become plans, and the best ones ship as pull requests.
Underneath: one database, one method (plan, pull request, reflect), one AI system. Not four apps.
Every Sunday ends the same way: run the week's check, which grades your proof and posts the scorecard. Run your PM agent reflection, which writes next week's plan. Record the demo. Log the hours the system now does without you.
Mandatory is synthetic, extra credit is real. The mandatory build runs on synthetic data you create yourself to represent your own company. Replacing a synthetic source with the real one is the extra credit, every week, and the only way to log hours saved. Your data stays in your own Supabase. Nobody at Edge8 sees it.
Not scored. The race is all hard weeks, so the install happens before the start line.
You don't prompt, you plan. A plan becomes a branch, a branch becomes a pull request, a pull request becomes a change on main. That loop is the whole job for the next seven weeks.
CLAUDE.md as the operating manual, plan mode, the plan file, review then build. Fork versus clone, main as the source of truth, the pull request with the plan linked, keeping your fork in sync with upstream.
Fork 8 Edges, live on your domain. Write your master plan, one section per office. Ship three small changes, each its own plan and its own pull request.
Three merged PRs, each linked to a plan. Site live. A history of branches and merges, never a push to main.
A hook that blocks commits containing secrets, and a first pull request back to 8 Edges.
What does your business know about itself. We never teach tables. The tables arrive as the answer.
Tables, rows, relationships, one table per concept, migrations in code, row level security on from day one. Structured versus unstructured.
Describe your business to Claude: who pays you, for what, how often, what a week produces. Extend the schema with your own entities, as migrations. Then generate your synthetic company to fit it: two years of data, a slow season, a customer who always pays late. Faking 400 customers convincingly is the data modelling.
Your migrations merged, every table secured, your customers on a screen by Tuesday, and the treasure hunt (questions only an organised database can answer) answered from your data.
Replace one synthetic source with the real one. An audit log on every write.
The road. Where the apps live, where the agents will live, and where every connection to the outside world runs.
Every push is a build, every branch a preview, main is production. Local connectors (on your laptop, for you) versus server side routines (on Vercel, for the company). Connect or sync, and why the central database wins.
Email and your meeting notes tool connected into the database as server side routines. Every meeting and every thread lands against the right person and company. First email out through Resend.
A new meeting appears in the database within a minute, attached to the right record. The sync run twice makes no duplicates.
A second system connected (your accounting or your CRM) with a sync that survives being killed mid run.
The business reads itself. Workboards and meetings come from the data. This week is as much about how your team runs meetings as about code, and we give you the script for that conversation.
Goals, rocks, workboards and meeting types in 8 Edges, and how to modify them for how you run the company, with EOS as the worked example. Admin interfaces and the design system.
Load Strategy to Execution. Customise the workboards and the meeting cadence. Run one real meeting from it. Wire your retreat prototype to the central database.
A teammate logged in and used it. One meeting run from the workboard, notes landing in the database. Loads on a phone.
A second meeting type (board, one on one, quarterly) with its own workboard.
Mass mailing is one message to everyone. Hyper personalised is the right message to each person, written from what the system already knows, including the meeting notes you connected in week 3.
Segments in SQL, the message framework, structured output per person, suppression and unsubscribe, deliverability, replies and clicks written back as activities.
One segment, one sequence. Every email generated from the record, the email history and the meeting notes. Sent through Resend. Replies and clicks flow back.
DMARC passes. A real send with no duplicates. Every click and reply is a row in the database. Unsubscribe works.
Two versions tested, measured on replies, the winner picked from the data.
The next thing starts from what the company knows, and from the people in it. Getting a team into a portal is harder than building it, so this week comes with the script too.
The team portal in 8 Edges: roles, who sees what, the home screen for a teammate. Ideas to plans: intake, triage, the plan file as the prompt, the pull request as the ship.
Build the portal for your team. The team submits ideas. Each idea becomes a plan. The best one ships as a pull request, built by agents in parallel.
Three teammates logged in. Five ideas submitted by people other than you. One idea shipped end to end: idea, plan, PR, live.
The portal on mobile, your team logging in from their phones, ideas voted on.
The AI hire. Planned as a folder, built in the cloud, onboarded like a human, with an approval queue as the boss.
Planning the agent as a folder: role, skills, context, routines, guardrails, definition of done. Building it in the cloud: a routine on Vercel that reads the database, calls Claude, does the job, writes back, on a schedule, every run logged, spend capped.
One AI team member with a real job: lead triage, meeting follow ups, or the Monday brief. Folder in the repo, routine on Vercel, actions behind an approval queue in your portal.
Ten test cases pass. It ran unattended on schedule. Every action traces to a logged run. The Monday brief arrived before you woke up.
A second agent in a different office, the two handing work to each other.
Nothing new. Everything from weeks 1 to 7 made production grade, documented, and shipped.
The access audit, tests on the critical journeys, backup and restore, mobile, monitoring. The runbook, so a human can keep the system running without you.
Run the full audit and fix what it finds. Write the runbook. Record the final demo: the whole system, four offices, three minutes. Report your hours saved graph.
Security advisor clean. Tests green. A restore from backup proven. Mobile score 90. Runbook in the repo.
A load test on the portal, and a restore proven by someone else on your team.
Any week, replace a synthetic source with the real one, pulled into your own Supabase. Points per source, capped per week. Hours saved can only be logged against real data, which is the point.
| Tier | Sources | Fits |
|---|---|---|
| Safe, yours alone | Your calendar, your sent mail, your meeting notes, LinkedIn connections export, the website lead form | Weeks 2, 3 |
| Business, low sensitivity | CRM contacts, newsletter list, accounting contacts, Google contacts, past event attendees | Weeks 2, 5 |
| Business, needs care | Invoices and payments, the support inbox, job applicants, proposals and contracts | Weeks 3, 6 |
| People data | Team roster, one on one notes, reviews | Weeks 4, 7, only after the team has been told |
The rules: the data stays in your own Supabase, nobody at Edge8 sees it, row level security is on before any import, and staff data only after the staff have been told.
| Component | How it scores |
|---|---|
| Weekly proof | Pass or fail, graded by a check you run against your own repo and database. Pass is required to count the week. |
| Extra credit | One harder cut per week, usually the real data version. Leaderboard points. |
| Hours saved | Logged every Sunday: hours a week the system now does without you. Scored against your own previous week, so everyone can win this one. |
| Demo votes | The cohort votes on the weekly demos. |
| Upstream contribution | A pull request accepted into 8 Edges, any week. Its own award. |
Ship day awards: highest total, most hours saved, best upstream contribution, best demo, and the comeback, for the biggest improvement from week 1.
Both cohorts run the same eight weeks at the same time, on one leaderboard. Entry is the week 0 gate: the stack installed and a one page essay on what your company should do differently. If you were at a retreat, the opt in is on your retreat portal.
What you get: eight deep dive videos, the official courses from Anthropic, GitHub, Vercel and Supabase mapped to each week, a prompt pack per build, Wednesday office hours, the weekly check that grades your work, and a cohort that ships beside you.