Impulsion Agentic Academy — One Brain, Many Arms

The big idea

What we're actually building

Imagine one shared "Chief of Staff" that never sleeps. It reads every client call (via Fathom), knows who signed what (via our agreements + Stripe), watches who's happy and who's drifting (via GoHighLevel), and can act — draft the deliverable, update the roster sheet, ping the right person. We don't each babystep it. It works on a schedule and leaves us one daily brief plus a short list of things waiting for a yes.

Analogy: Think of it as hiring a brilliant, tireless assistant who has read everything we've ever done — but who always asks permission before spending money or messaging a client. We give it arms (the APIs) and a conscience (the safety gate).

120

DFY clients by end of July

Pods of 4

meet every other week, 1 hour

$697+/mo

floor price — and we raise it

One brain

many arms, customizable per person

Two teams, one brain. Adam & Brian build the apps that ship unbelievable deliverables. Meg, Ranjeet, Farah, Kevin & Dabeer present the solutions and service the clients into momentum pods.

Vocabulary that matters

Token vs. API vs. MCP

These three get mixed up constantly. Here's the clean mental model.

🔑 Token / API Key

A password for software. It proves "this program is allowed to act as us." Looks like sk_live_… or cfut_….

Analogy: a hotel key card. It opens certain doors, for a certain time, and can be deactivated instantly.

🔌 API

The set of doors a service offers. "Stripe's API" = all the things you're allowed to ask Stripe to do (list charges, refund, etc.). The token decides which doors your card opens.

Analogy: the building with labeled doors. The API is the building; the token is your card.

🧠 MCP

Model Context Protocol — a universal adapter that lets the AI use many APIs through one standard plug, so we don't hand-wire each one. It also lets the AI use tools safely and consistently.

Analogy: a universal travel adapter — one plug, many countries' sockets.

In one sentence The API is what a service can do, the token is your permission to do some of it, and MCP is the standard way our AI plugs into many of them at once.

Core distinction

What's the difference between a Skill and an Agent?

📜 A Skill

A written playbook the AI loads when a task matches. It's knowledge + instructions: "here's exactly how we write a follow-up email," "here's our refund policy." Skills don't do anything by themselves — they make the AI better at a specific job.

Think: an SOP in a binder the assistant grabs off the shelf.

🤖 An Agent

An actor with a goal, tools, and autonomy. It can take steps, call APIs, make decisions, and loop until done. An agent uses skills and tools to accomplish something.

Think: the assistant themselves — who reads the SOP, then goes and does the work.

Analogy: A skill is a recipe; an agent is the chef. The chef (agent) can read many recipes (skills), use the kitchen tools (APIs), and actually cook the meal. A recipe alone never cooks anything.

Where we work

VS Code vs. Claude Code (vs. the Desktop app)

Same underlying brain (Claude). Very different cockpits.

Surface	What it is	Best for	Watch-outs
Claude Desktop app	Chat window. Great conversation, can use MCP tools you connect.	Thinking, drafting, asking questions, light tool use.	No real file/project workspace; not built for shipping code.
Claude Code	A terminal-based agent that reads/edits your whole project and runs commands.	Power users who live in the terminal; scripted agent runs.	Terminal-first; less visual; easier to fire broad actions fast.
VS Code + Claude (us)	A full editor where the agent sees your files, runs commands, deploys, and you see every change visually with undo.	Building + deploying real things as a team, with guardrails and version history.	Slight learning curve — but it's the safest, most visible way to build.

Why we use VS Code You see what the agent changed, you have undo / git history, you can review before deploy, and the same window builds and ships. This page was built and deployed live from VS Code — that's the whole point.

The Claude family

Claude in Chrome, Desktop, Cowork, and Design — what's the difference?

🌐 Claude in Chrome

A browser extension where Claude can see and act on the web page you're on — fill forms, click, summarize, pull data from a dashboard.

Great for: working inside GHL, Stripe, or any web tool alongside you.

💻 Claude Desktop

The standalone chat app on your Mac. Connect MCP tools, talk through problems, draft. Your everyday thinking partner.

Great for: ideation, writing, quick tool calls.

👥 Claude Cowork

The collaborative/agentic workspace where Claude works on longer multi-step tasks alongside a team, with shared context.

Great for: a shared agent the pod can all direct.

🎨 Claude Design / Artifacts

Where Claude produces visual, interactive deliverables — pages, mockups, mini-apps you can see and share immediately.

Great for: client-facing deliverables that need to look great fast.

⚠️ The Claude product line moves fast and names/availability change. Treat this as the working picture as of June 2026 — we'll confirm specifics per surface as we adopt each one.

The arms

Which integrations help us — and what each one unlocks

Integration	What the brain can do with it	Status
Fathom	Read every client call + transcript; summarize, extract decisions, spot churn signals.	✅ live
GoHighLevel (GHL)	See pipelines, appointments, conversations; trigger notifications when a client is angry / quiet / thrilled.	⏳ to wire
Stripe	See who paid, who churned, MRR; cross-reference past clients for reactivation.	✅ key tested
Cloudflare	Publish pages, dashboards, and deliverables to live URLs (like this page).	✅ deploying
Google (Docs / Sheets / Drive)	Read client notes; keep the shared client roster sheet updated in real time.	⏳ to wire
Zoom	Pull recordings + transcripts for calls not on Fathom.	⚠️ needs S2S app
Basecamp	Project/task visibility so deliverables route to the right owner.	⏳ to wire
Google Chat	Push the daily brief + alerts where the team already talks.	⏳ to wire
Twilio	SMS notifications/outreach (after A2P compliance).	⚠️ needs Account SID

VS Code extensions worth having: the Claude/Copilot extension, GitLens (history), REST Client (test APIs), and the Cloudflare/Wrangler CLI for deploys.

Read this twice

Security, Blast Radius & Client Data

Power cuts both ways. An agent that can update GHL can also mess up GHL. Here's how we stay safe.

What's the worst that could happen?

Worst cases

A leaked key lets someone drain Stripe or read every client's data.
An over-eager agent messages the wrong client or mass-texts a list.
Private info (a tax return, an SSN) gets pasted into a client-facing draft.
A bad command deletes data with no undo.

How we protect

Least privilege: keys are scoped to the minimum (read-only where possible, restricted keys for money).
Blast radius: separate keys per service & per client, so one leak ≠ everything.
The wall: private/personal data lives in a vault no agent can reach.
Approve-before-act: money, messages, and irreversible actions wait for a human "yes."
Never commit secrets; rotate keys on a schedule; encrypt at rest.

Blast radius, explained: if a key is a hotel key card, "blast radius" is how many doors that card opens. We hand out cards that open one room, not the master key. So if a card is lost, the damage is contained to one room.

The golden rules (everyone)

Never paste a real key into a chat, email, Slack, or a public file. Keys go in the vault.
If you ever think a key leaked: rotate first (revoke + regenerate), investigate second.
Treat client data as borrowed, not owned. It never leaves where it belongs.
When an agent offers to do something irreversible, read the proposal before you approve.

How ours is wired

One brain, many arms — the layout

🧠 The Brain

The shared, curated knowledge + the Chief-of-Staff orchestrator. Knows our playbooks; never holds your private life.

💪 The Arms

The integrations: Fathom, GHL, Stripe, Google, Cloudflare. Each one scoped and labeled.

🚪 The Gate

Plain-code safety (not AI): invariants, permission tiers, kill switch, and the "needs Brian's yes" queue.

The leverage point Each of us gets an arm of the same brain, customized to our role — and we tune it iteratively. You don't manage the agent step-by-step; you read its brief and approve what matters.

Do this week

Your assignment

Small, concrete, and designed so you learn by doing. ~30–45 minutes.

Adam & Brian — the Builders

Stand up the first app/deliverable arm in VS Code.
Wire the GHL + Stripe reads (least-privilege keys) to draft a client status view.
Ship one "unbelievable deliverable" prototype to a live URL.

Meg, Ranjeet, Farah, Kevin & Dabeer — the Service Pod

Make sure every client meeting records to Fathom and lands in the shared folder.
Drop your client Google Docs notes into the shared Drive folder beside the roster.
Open VS Code once, ask the brain one question about a client, and read its answer.

Everyone — the 4 steps

Use Fathom on every client call. No exceptions — it's the brain's eyes and ears.

Put notes + recordings in the shared folders. One place, so the whole team and the brain can see them.

Read the daily brief. Five minutes. Approve/deny what's waiting for a yes.

Tell us how you'd customize your arm. What would you want it to do for your clients? We build it.

The "why" behind the goal: 120 clients, pods of 4, $697+/mo

Pods of four that meet every other week for an hour create accountability and community — happier clients, lower churn, easier delivery. The brain helps us (a) reactivate past believers who stalled, (b) deliver consistently, and (c) spot churn early. The price floor is $697/mo and we intend to raise it as the deliverables get more "unbelievable."