Phase 1

The Factory with a New Motor

We bolted the motor on. Let's see what it can do.

How the Machine Works

A large language model is a next-token prediction machine:

1

Tokenization

Text is broken into small pieces (“tokens”, roughly 3/4 of a word).

2

Pattern Recognition

Trained on massive text, it learned statistical patterns about which tokens follow others.

3

Probability

Given a sequence, it predicts the most likely next token. Repeats. That's it.

4

Emergent Capabilities

At scale, surprising abilities arise: reasoning, coding, analysis. It doesn’t “know” - but the results can be remarkable.

See It in Action

Watch how tokens flow through the model - from input to prediction, one token at a time.

Why AI “Forgets” & Can't Go Back

Stateless

Every conversation starts from zero. The AI has no memory of yesterday's chat. It's not being difficult - it literally doesn't know.

Autoregressive

Tokens are generated left-to-right, one at a time. It can't revise earlier tokens once written. That's why your feedback matters - the AI literally can't go back and fix it on its own.

This is not a flaw - it's the architecture. Pattern matching at scale, not reasoning. The results can be brilliant, but it's statistics all the way down.

The Full Pipeline

From your prompt to the AI's response - this is what happens under the hood.

Not All Models Are Equal

Rule of thumb: bigger ≠ better. Match the model to the task.

ModelContextPrice / 1M tokensBest For
Claude Opus 4.61M$5.00Complex analysis, strategy
Claude Sonnet 4.61M$3.00Day-to-day work, coding
GPT-5.4272K–1M$2.50Complex reasoning
Gemini 3 Pro1M~$2.00Huge document analysis
Llama 4 Scout10M$0.12–$0.88Privacy, open source
DeepSeek R164K$0.55Budget reasoning

Watch out for thinking overhead: Reasoning models burn tokens on simple tasks - thinking thinking thinking… “hello.” An “expensive” model isn’t always better. Match complexity to the task.

The #1 Rule: Context Is Everything

Don't rely on AI “knowing” - it doesn't. Research first. Always. Ground it in reality.

Bad

“Summarize this project”

Good

“You are reviewing a pharma supply chain audit for a Belgian food producer. Here is the full report: [...]”

The difference between a useless answer and a brilliant one is context.

The “New Hire on First Day” Analogy

You wouldn't dump every company document on a new hire's desk and say “make me an invoice.” You'd extract the relevant info and write a clear brief. Treat AI the same way. This connects directly to the instruction documents section later.

Developer Productivity: GitHub Copilot

0%

Faster task completion

0x

Faster PR cycle

0%

Code written by AI

0%

AI code kept in final version

90% of Fortune 100 companies use Copilot. 20 million users by mid-2025. Caveat: AI-assisted code can increase issue counts (~1.7x) if not paired with proper review. Speed without governance = faster mistakes.

Sources: GitHub Research 2022 · GitHub × Accenture 2024 · GitHub Economic Impact 2023 · CodeRabbit Report 2025

Demo

GitHub Copilot in action

Writing a function with Copilot autocomplete, tab-accepting suggestions, and inline chat to refactor code.

Generating Images

ToolBest For
MidjourneyBest aesthetics/art
GPT ImageEasiest, best prompt adherence
Adobe FireflySafest commercially (licensed data)
Stable DiffusionFull control, open source
FLUXPhotorealistic, fast

Note: Claude does NOT generate images - it excels at image analysis/vision only.

The Ghibli Story (March 2025)

GPT-4o went viral with Studio Ghibli-style images. Miyazaki had called AI animation “an insult to life itself.” Sam Altman: “our GPUs are melting.” Major copyright debate - style is NOT explicitly protected by copyright.

Demo

Image generation

Generating an image with GPT and/or Midjourney, comparing outputs and styles.

Gamma & Presentation Tools

Gamma.app - AI-powered presentations. Agent researches, generates, and restyles entire decks via conversation. Good for quick first drafts - always review and refine.

Demo

Gamma presentation

Creating a presentation from a single prompt in Gamma, then restyling it via conversation.

Markdown: Why and How

Universal & Portable

Every AI tool reads/writes it natively. Works in any editor, converts to HTML/PDF/slides.

Structured

Headings, lists, tables, code blocks - AI understands the hierarchy.

Version-friendly

Plain text = easy to diff, track changes, and collaborate.

Lightweight

Tiny files compared to .docx or .pptx - easy to share and store.

Quick syntax

# Heading 1
## Heading 2
- Bullet point
**Bold** and *italic*
| Table | Header |
[Link](url)
Try it live: Markdown Editor

Google AI Studio

Free1,000 daily requests across all Gemini models
Vibe CodingDescribe an app → get a working app → deploy to Cloud Run or Netlify
localStorage5–10MB persistence for personal apps, no backend needed
Context2M context window (Gemini 3 Pro) - entire codebases or document sets
System InstructionsPersistent context for consistent behavior
Demo

Google AI Studio vibe coding

Describing a simple CRM app → AI generates it → working app → deploy to Cloud Run or Netlify. All free, no code.

Data stays local: Google AI Studio creates the app with AI, but your actual data (CSVs, spreadsheets) is processed in the generated app on your device - not sent to the AI model.

Better context = better apps. Just like prompting, the more specific your description, the better the result. Research first, describe clearly.

Ask AI to Help You Ask Better Questions

When you need advice from an expert, you don't walk in and ask random questions. You ask someone to help you figure out what to ask.

The same applies to AI. Before asking a complex question, ask the AI: “What should I ask you to get the best answer about X?”

Direct (often weak)

“How do I improve my team's productivity?”

Meta-question (better)

“I manage a 12-person finance team. What questions should I ask you to get actionable advice on improving our quarterly close process?”

This is a teaser - in Phase 2, we'll turn this into a full meta-prompting workflow.

Demo

Meta-questioning demo

Asking AI 'what should I ask you?' and comparing the quality of answers to direct questions.

Instruction Documents & Custom Instructions

You wouldn't hire someone and explain the job differently every morning.

Claude: Projects with custom instructions / system prompts

ChatGPT: Custom GPTs / “Customize ChatGPT” instructions

Google AI Studio: Persistent system instructions

Think of it as an onboarding document for your AI. The better the document, the better every interaction.

✍️

Write your first instruction document

  1. 1

    Open Claude (claude.ai) or ChatGPT (chatgpt.com) on your laptop/phone

  2. 2

    Pick a task you do repeatedly at work (e.g. meeting summary, email reply, report)

  3. 3

    Ask the AI to write you a reusable instruction prompt:

    Write me a reusable instruction prompt for [your task]. Include placeholders for the parts that change each time.
  4. 4

    Test it: paste in real (or example) content and run the prompt

  5. 5

    Save it somewhere you'll find it tomorrow

So... is this enough?

We bolted on the motor. Things go faster. But look at the numbers again: 88% of companies are here. Only 6% are seeing real impact.

The factory still has the same belts, the same layout, the same bottlenecks. The email still sits in the mailbox for a day.

Let's rip out the belts →