NYU Leslie Entrepreneurial Institute Spring 2026 AI Systems · Program Design

Startup Intelligence System: A Second Brain for Founders, Adopted by an Entire Cohort in Two Weeks

I built a shared AI discovery partner for 18 founder teams and rolled it out in two weeks. It held every customer interview, knew every team's business, and coached them on their own evidence. Founders kept it. Coaches rebuilt their workshops around it. I presented it at the GCEC annual conference in October 2026.

18 teams ~900 interviews in the Spring 2026 Startup Sprint, about 50 per team
25 of 36 Surveyed founders will keep using it. Zero plan to stop.
1 in 3 Teams not admitted to the next program kept using it on their own
Best Use of AI An award normally reserved for participating teams

The Problem

Nobody holds 50 interviews in their head. By week two, founders forget what early customers said. Coaches only hear what founders choose to report. By the time coaching happens, the signal has been interpreted twice.

Out-of-the-box AI made it worse. Founders were already using ChatGPT and Claude, but every chat started from zero. They spent their time re-explaining the venture instead of learning from it.

We needed an AI that already knew every team: the business, the method, and what the team learned yesterday.

"We really need to create a centralized mind within our own team just so we can really be on the same page with our understanding of the problem and therefore our vision of the solution."

— Founder, Agora

Before and After

Before
Manual coaching check-ins
No persistent AI support between sessions
No structured evidence tracking
Teams rebuilt context from scratch every session
Workshops captured on flipcharts and phone photos
After
Every team has a persistent AI thinking partner that reads their interviews
Assumptions tracked against a four-level evidence framework
Workshops designed and run AI-first
Coaches working inside each team's workspace
Context that compounds across the full two weeks
Coaching check-ins start from cited quotes, not recall

The Hardest Design Problem

The Spread

The hardest problem here was never the AI. It was the range of people who had to use it.

The cohort ran from founders who had been building AI products for years to founders who had never opened a terminal. A system that only works for the top of that distribution is a perk for the already capable. So I set the least technical founder as the design target and let everyone else inherit the same system.

Three decisions followed from that one constraint.

Two parallel tracks on a single methodology

Technical founders run Claude Code in the terminal, connected to Notion over MCP. Everyone else uses Claude in the browser with the Notion connector. Same workspace, same skills, same outputs, different door in.

Setup stripped to three steps

Duplicate a Notion template, drop in a config file, run one command. Nothing else stands between a founder and a working system.

A silent start protocol

Claude reads the team's context, initializes it if empty, and installs any missing skills before the founder types a word. A founder's first interaction is a working system, not an empty prompt box.

The result was near-universal adoption across the cohort with almost no technical issues. For a two-week program where every hour counts, the reliability was a design outcome, not an accident.

How It Works

Two layers, ten skills, zero happy ears

01
Notion stores. Claude analyzes.

Each team gets a Notion workspace: the Discovery Records Manager with every interview, a versioned Business Model Canvas, workshop outputs, the written methodology, and a Context Log. Claude reads it through MCP. Storage and reasoning stay separate, so every answer traces back to a source.

02
The method as skills

Ten custom skills turn the Institute's discovery methodology into one-command workflows. /discovery-ladder maps a team's evidence for one customer and one problem. /customer-journey drafts a journey from the team's own interviews. /interview-feedback shows founders, quote by quote, where they led the witness or started pitching. Skills live as Notion pages that Claude installs, so one update reaches every team.

03
Built to resist happy ears

The AI never tells founders what to do. It never validates, never celebrates, and never offers next steps unprompted. It surfaces evidence and flags contradictions, the same way our coaches coach instead of consult.

The methodology itself, the Discovery Ladder and its scoring rubrics, was developed by my colleague at the Institute. My job was to build the system that runs it.

Tools and Stack

What I built it with

Claude Code
Terminal access for technical founders. Full skill support and persistent project context.
Claude.ai
Browser access for everyone else. Notion connected as an integration, skills used as prompt workflows.
Notion
Persistent workspace, interview database, session memory, and the read-only methodology library.
Notion MCP
The live read and write bridge between Claude and each team's workspace.
Custom skill system
Ten installable commands that turn the discovery methodology into one-command workflows.
CLAUDE.md
A dedicated instruction file per team, so every founder's Claude behaves the same way.

What I Actually Built

The software was the easy part. The work was adoption.

I built, tested, and rolled the system out in about two weeks. It was live on day one of the sprint. I trained all 50 founders and every coach myself. The cohort ran from founders who had built AI products for years to founders who had never opened a terminal. I designed for the second group.

This was the second AI system I took to full adoption at the Institute. The first was the Leslie coaching copilot. The lesson carried over: skeptics become advocates when the system makes their own work better, not when you tell them it will.

Results

We surveyed founders at the end of the sprint. 36 responded.

The interview record won. Founders named it the most helpful feature, more than any skill.
25 of 36 will keep using it. 11 are unsure. Zero plan to stop.
Coaching moved from recall to evidence. Coaches stopped spending the first 20 minutes of a check-in reconstructing what founders heard. They started from cited quotes.
Workshops went AI-first. Claude drafted each team's customer journey and assumptions from their own interviews. Workshop time went to debating which assumptions were riskiest.
Teammates self-served. One teammate ran the interviews. The others read them in Notion instead of waiting to be briefed.
It reached past the program. About a third of teams not admitted to the summer program kept using it. Four teams rethought their core customer on day one, including one that had reported more than 100 interviews.
It may predict team quality. System scores matched our coaches' hand-scored top two teams. Early, but promising.

One team shows what the record is for. Counterpart, an app for actors, checked claims from an industry meeting against their own interviews. 18 of 50 customers had independently mentioned a fear of AI. They repositioned. An ad written in their customers' own words drew about 60 trial signups.

"It called me out on my initial reluctance to dig in when I heard people identify pain points."

— Founder, SoundMesh

What Broke

I ran this as an experiment, so here is what failed.

AI inside the storage layer. Notion AI's autofill flipped classifications on re-runs and invented quotes. More than 10 teams named it. Once founders saw one field flip, they stopped trusting all of them. Analysis moved to Claude.
18 copies of everything. Claude skills updated across teams. Notion AI prompts did not. Every prompt fix meant 18 manual edits.
No context, no output. Teams that started with empty workspaces spent week one feeding the machine. We needed twice the onboarding.
Our best skill was hidden. /interview-feedback was the most loved feature and the least discovered. It now runs on every interview.

Trust came down to two questions: does it cite its sources, and does it behave the same way twice?

What's Next

The pilot ran on Notion and Claude, configured by hand. That works for about 10 to 20 teams. It breaks at scale. I'm now leading the move to a proper application that scales across our programs, starting with the open Bootcamp funnel of 600 to 800 teams. The same record serves us as operators: we can select teams on their evidence instead of their pitch.

Design Principles

01 Design for the floor, not the ceiling. Setting the least prepared user as the target is what produced universal adoption. A system built for the power user only ever serves the power user.
02 Make every output self-contained. A founder who has never read the framework can act on the result, because the insight travels with it and references nothing external.
03 Reflect evidence, stay out of the way. The system surfaces contradictions and cites its sources. It never tells a founder what to do next unless asked. Under time pressure, unsolicited advice reads as noise.
04 Build the method and the tool in the same loop. The methodology was written into the system while the system was being built and used. Ship, learn, revise, ship again, the same loop we ask founders to run.