NYU Leslie Entrepreneurial Institute Fall 2026 AI Systems · Operations

Leslie OS: Giving an Entire Team the Same AI Teammate

Every teammate at the Institute runs their own Claude. With Leslie OS, every one of those Claudes already knows how the Institute works: our programs, our people, our tools, and our conventions. Nobody configures AI from scratch, and nobody gets a version that guesses.

Entire team Adopted across the Institute
12 skills 8 core skills released, 4 in review
6 tools Connected: Notion, Slack, Granola, Gmail, Google Calendar, Google Drive
~10 min To onboard a new teammate, with no enterprise plan

The Problem

Out of the box, an AI assistant knows nothing about your organization. Ask it to look up a founder, log a coaching note, or build an on-brand deck. It either can't, or it invents an answer.

The usual fix is that everyone teaches their own assistant. That fails twice. Each person repeats the same work. And every assistant drifts in a different direction.

We had already built two AI systems for founders. Our own team was still working with AI that didn't know us.

Before and After

Before
Each teammate set up AI on their own, or not at all
Facts re-explained in every chat
Decks and copy checked for brand by hand
Answers that guessed at programs, people, and rules
After
One shared setup every teammate inherits
Programs, people, and contacts fetched live from Notion
Brand, voice, and deck templates applied by skills
New teammates productive in about 10 minutes

How It Works

Three layers, sorted by how often they change

The design rests on one rule: each layer holds the least-changing version of what it needs, and upper layers point to lower ones instead of copying them.

01
Always-on instructions

A short block each teammate pastes into their Claude settings. It says who we are, names the baseline tools, and tells Claude to load the master skill for Institute work. It loads in every session, so it stays tiny.

02
The master skill

A skill called leslie-os loads only when the work is Institute-related. It orients Claude and routes to the right specialist skill. It points to Notion for facts instead of restating them, so it rarely needs reinstalling.

03
The Notion knowledge base

The live source of truth: programs, people, contacts, and conventions. Claude fetches it at runtime, so it is never stale.

Why the split: without an enterprise plan, nothing can be pushed to every account centrally. So the layer every person maintains is the smallest one, and everything that churns lives in Notion.

Capabilities

What it does today

Institute data
Look up or update a company, founder, application, review, or cohort record, following the workspace's write rules.
Meetings and coaching
Pull notes, decisions, and action items from Granola, and recognize coaching sessions.
Scheduling
Check the program calendar, book eLab rooms, and find time across teammates.
Files
Find and file documents in the shared Drive by its naming conventions.
Brand and production
Build decks on the Institute template and check copy against the style guide.
People and comms
Find a teammate's Slack handle, coaching link, or Zoom link, and draft messages.

Tools and Stack

What I built it with

Claude (Cowork)
Every teammate's AI, running the same instructions and skills.
Notion
The knowledge base, the skill library, and the onboarding checklist.
Six connectors
Notion, Slack, Granola, Gmail, Google Calendar, Google Drive.
Skill library
12 skills with a tier (Core or Shared) and a status (Draft, Reviewed, Released).

What I Actually Built

I designed the architecture and did the integration. This was a team build: other colleagues built several of the skills. My job was to make the parts fit and keep them from drifting.

The hard part was governance, not prompts. Every knowledge page and every skill has an owner and a review cadence. Skills move from Draft to Reviewed to Released. And one rule gets revisited more than any other: facts live in exactly one place.

This is the third AI system I have taken to full adoption at the Institute, after the coaching copilot and the Startup Intelligence System. The first two served founders. This one serves the team.

Where It Stands

The core is live: the knowledge base, the six connectors, and eight released core skills. Four skills are still in draft: researcher decks, the coaching playbook, new-cohort setup, and a cross-system data ontology under review. The team keeps filling in the knowledge base.

I don't have impact numbers yet. Adoption came first. Measurement comes next.

Design Principles

01 Layer by change rate. Tiny and stable in the always-on block, procedure in skills, anything that churns in live Notion.
02 Point, don't copy. Every duplicated fact is a future contradiction. The project caught its own drift this way more than once.
03 Skills are thin procedure over live data. A skill stays small because the facts it needs are fetched, not baked in.
04 Guidance, not enforcement. Skills guide Claude. They don't block it. So the guardrails sit where they matter, such as a confirmation before anything is sent.