01-projects/mammoth-growth/2026-05-13-ai-demo-day

01 script v1

2026-05-10·project·status: draft-v1·! low (script only; demo data is real but PII-safe per prep-plan)
mammoth-growthdemoscript-draftray-starter-kit

Mammoth Growth Demo — Script v1

Voice notes for delivery

Cue legend

[SLIDE: X] — slide change [DEMO: Y] — live action on screen [PAUSE] — beat for emphasis (NN-NN) — running clock (target)

(0:00-0:30) Hook

[SLIDE 1: title]

I have an AI COO. He runs on a Mac mini in my bedroom. He woke up at 6:30 this morning, wrote me a brief, processed last night's incoming newsletters, ran a vault health check, and is currently watching for the next Notion task to come in. I haven't asked him to do any of that today.

[PAUSE]

What I want to show you in the next nine minutes is what's possible when you stop using AI as a thing you invoke for one-off tasks and start running it as a teammate that earns its rules from your failures.

Let me show you what that actually looks like in practice. We'll start with the morning.

(0:30-2:00) Live morning brief

[DEMO: switch projector to iPhone cast, Messages app open to today's thread]

This is iMessage. The thread you're seeing is between me and Ray — that's what I named him. Top message landed at 6:30am.

[DEMO: scroll to today's morning brief, give them 5 seconds to read]

Notice the structure. Open decisions at the top with one-tap links into HQ. An energy-check line — I reply with a number 1-10 and Ray adjusts his queue depth for the day, fewer strategic asks if I'm at 3, more if I'm at 8. Overnight research links if anything fresh landed at 1am. Today's bets, today's calendar, suggested focus.

[DEMO: tap the decisions link]

Tap a link, jump straight into HQ — that's where Ray's whole working state lives. Let me show you that next.

(2:00-4:00) Live HQ tour

[DEMO: switch projector back to laptop, browser already open to hq.raydata.co]

This is HQ. It's an Astro site that pulls from a vault on my Mac mini. Three big surfaces.

[DEMO: click /vault]

/vault is the nervous system. Every concept I've ever filed lives here. Cross-linked. Queryable. Let me open one I filed this morning.

[DEMO: click into /vault/06-reference/concepts/2026-05-10-ray-architecture-introspection]

Eight layers of capability. Thirteen unhobbling moments. Five composability chains. I asked Ray to do this introspection today and he wrote it. Now it's a permanent surface I can reference every time I need to think about how the system actually works.

[DEMO: back, click /decisions]

/decisions is the rail. Every decision I make, big or small, gets a page like this. Title, framing, options, status. Open ones at the top.

[DEMO: open one recent decision page on the projector]

Read the framing. See the options. Pick one. Watch what happens next — this is the part most people haven't seen before.

(4:00-6:00) Click-back demo (the wow moment)

[DEMO: pull out iPhone, switch projector to iPhone cast, open the same staged demo decision page]

Same page on iPhone now. I'm going to tap one of these option chips.

[DEMO: tap "Option A" or whatever the staged option is]

Watch what just happened. Messages app opened. Pre-filled with a structured payload that tells Ray which decision, which option, and any context I want to add.

[DEMO: hit send]

[DEMO: switch projector back to laptop, scroll the Ray session live so they can see the iMessage land]

Real-time. Ray just received it. He's parsing the structured payload. Writing the decision-log file to vault. Closing the corresponding Notion task. Done.

[PAUSE]

This is generative UI as a return channel. One tap on my phone, one decision logged forever, and Ray now knows what I decided without me typing a single character of context.

If you remember nothing else from today — this pattern is reproducible. You build the HTML page, you encode the form-state into an sms: URL, and the iMessage MCP catches the structured input on the other side. The whole stack is browser primitives plus iOS plus a local agent. No new infrastructure.

Let me show you the part that's actually the moat.

(6:00-8:00) The harness ratchet (the depth moment)

[SLIDE: CLAUDE.md hard rules visible on screen]

This file is called CLAUDE.md. Four rules. That's it. These are everything Ray needs to know about the kind of mistakes I make.

[PAUSE]

Each one was earned. Rule 1, always check the date — earned because Ray once told me it was Tuesday for two days straight. Rule 2, channel responses must go through the reply tool — earned because for a week I thought Ray was ignoring my messages, and actually he was just writing replies into a session log only I could see. Rule 4, route any artifact over 5KB through a subagent — earned because reading newsletter HTML into the parent session was rotting his context for the rest of the day.

[SLIDE: switch to a memory file, show feedback_calibrate_overconfidence.md]

This is one of about forty memory files. Every founder correction, every observed failure, gets one. They get loaded automatically into every session.

[SLIDE: switch to today's session showing a system-reminder block]

Last week, Ray was overconfident on a recommendation. I caught it. He wrote it as a memory file. This morning, that memory file showed up as a system-reminder before every recommendation he made.

[PAUSE]

I never re-trained the model. I tightened the harness. The system gets better at being mine, week over week, because every failure becomes a permanent rule.

This is the part most teams are leaving on the table. Everyone is racing to upgrade models. The teams getting outsized returns are the ones running this discipline — every mistake earns a rule, every rule survives forever, and the harness compounds while the model stays the same.

[PAUSE]

So here's the parting gift.

(8:00-9:30) Parting gift

[SLIDE: starter-kit page with repo URL prominent]

I'm leaving you all the recipe.

[DEMO: open the GitHub repo on the projector]

Public repo. Inside: a SOUL template, a CLAUDE template with the four universal hard rules, a vault folder scaffold, twelve starter skills — the ones that were the actual unhobbling moments for me. The audit script. The cron schedule. A six-week onboarding ratchet doc that walks you through Day 1 to Week 6.

[PAUSE]

Six weeks from now, you have your version of this. Different from mine — your CLAUDE.md will look different, because your specific failures will be different. But the discipline is the same.

Three things to do Monday.

[SLIDE: 3-step action list]

One — install Claude Code. Two — write a three-line SOUL.md describing who you are and what kind of decisions you want help with. Three — pick one recurring task in your life — your morning email triage, your weekly report, your reading queue, anything that comes back every week — and wire it to one channel.

The ratchet starts there.

(9:30-10:00) Close

[PAUSE, look around the room]

I'm here for questions. But if there's only one thing you take from today: stop invoking AI for one-off tasks. Start running it as a teammate that earns its rules from your failures.

That's how you go from "I have a tool I use sometimes" to "I have a teammate that runs my life with me."

[SLIDE: Q&A holding slide with repo URL persistent in footer]

Thank you.

Word count check

Total: ~1,140 words. At 130 words/min spoken (slow conference pace), that's 8m 45s — gives 1m 15s buffer for live demos taking longer than expected.

Open script-level questions for founder

  1. Repo name in the parting-gift slide — final call?
  2. The "Mac mini in my bedroom" line — keep self-deprecating, or shift to "in my home office"?
  3. The CLAUDE.md hard rules section — do I name all 4 with their failure stories, or stick to 3 examples for time?
  4. Q&A slot — handled inside the 10 min, or audience expects to hold questions and the 10 min is monologue-only?

Risk checklist (Tue rehearsal items)

Related