← → navigate · space next · n presenter notes
The Baker's Dozen · Workshop 1 · Session 1

Cultivated AI

Grow, don't Chase.
01 · Welcome
About me

Shalin Shodhan

20 years in games & film.

  • Pixar Animation Studios 6 feature films · 3 Academy Awards
  • Electronic Arts · Spore BAFTA Award for Technical Achievement
  • Lumosity Principal Game Developer
  • Masala Games · Founder Detective Dotson · Word Mess (#1 in 21 countries)
  • Carnegie Mellon University MET · Entertainment Technology
Shalin Shodhan
shalin@masalagames.com
Detective Dotson Toy Story 3 Cars 2 Monsters University Inside Out Finding Dory Spore Word Mess
02 · About

Two sessions.
Two hours each.

today
01

How it works, and what's out there

Why it matters, how it behaves, which tool for what. Half of it hands-on.

02

Agents, skills and a plan

Build a browser extension, write a skill of your own, start your 90-day plan.

03 · Two sessions
Before anything else

Relax. It's not here for your job.

Teach it the routine parts. Spend the time you get back on the parts that need you.

The chores you said you'd hand over first
Keeping spreadsheets updated18
Meeting notes & follow-ups14
Invoices & paperwork12
Checking competitor prices9
The weekly sales report9
from your questionnaire · 30 answers
04 · Relax
Principle 13 · The human edge

Conviction. Art. Sense. Judgment.

These come from you. Not the machine.

05 · Edge
Principle 15 · A mental model

Every AI task is one of three things.

Know

Information

Research · reasoning · brainstorming. Input and output are understanding.

e.g. Deep Research Claude

Do

Transaction

Send an email · make a purchase · book, pay, post. Input: intent. Output: a transaction of some sort.

Delegate these with the most care. A bad draft can be fixed. A bad transaction already happened.

beware Unsupervised agents OpenClaw

Show

Artifact

Decks · sites · videos · software. Input: a message. Output: something consumable.

e.g. Claude NotebookLM Nano Banana Seedance
06 · Know · Do · Show
The path this workshop climbs

The climb.

Quoter
14 of you
→
Curator
16 of you
→
Author
0 · the 90-day goal
→
Orchestrator
0 · not yet needed
from your questionnaire · 30 answers where do you sit today?
07 · The climb
How it actually works

It began as a sketch of the brain.

Neurons vs neural networks: a single neuron, connected neurons, and a multi-layer artificial neural network
08 · Brain
How it actually works

The learning lives in the weights.

Every connection is a number. Training nudges trillions of them until the guesses get good. The finished model is that file of numbers.

09 · Weights
How it actually works

It predicts the next word. Nothing more.

10 · Next word
Why it behaves the way it does
11 · What follows
Laptops open →

Lab 01 · Context Explorer

talkingtomachines.xyz/tools/context
12 · Hands-on
Nine of you asked

Where does my data go?

To the model maker's servers, under your plan's rules. Free plans may learn from your chats. Business plans promise not to. Every tool has a toggle.

nothing you wouldn't email an outside vendor
13 · Your data
Back to the screen →

The lay of the land.

Same brief · three machines · live

14 · Break
Who's good at what

The wider map.

LLM

The daily drivers, plus the open ones from China.

Claude ChatGPT Gemini DeepSeek Qwen Kimi

Images

Product shots, edits, concepts.

Nano Banana ChatGPT Image

Video

Clips from a sentence.

Seedance 2.5 Kling 3

Music & audio

Songs and voices from a brief.

Suno ElevenLabs

Product photography & branding

Pack shots, backgrounds, brand looks.

Photoroom Pomelli

Research

Your documents and the open web, with citations.

Gemini Notebook Perplexity

Software agents

Describe it; they build it. Session 2 lives here.

Claude Code Codex
15 · The map
The lay of the land · one more thing

Open weights: models you can own.

Remember the file of numbers? Some labs publish it. Download it, run it on your own machine. Nothing leaves the building.

Llama · Meta DeepSeek Qwen · Alibaba Gemma · Google Mistral
16 · Open weights
Staying current without chasing

Grow, don't chase.

Tools will keep changing under us. What holds: your principles, your examples, your saved sessions.

one team conversation a month
17 · What holds
Principle 01 · Rate of improvement

Learn for the AI of tomorrow.

Build around approach, not tools.

18 · Rate
Principle 02 · The real problem

Getting the most out of AI is not a technical problem.

It's a communication problem.

19 · Communication
Principle 03 · The core loop

Three moves. On repeat.

Choose
→
Communicate
→
Evaluate
20 · The loop
Choose · Principle 04

Pick the right AI for the task.

When in doubt, ask AI which AI to use.

21 · Choose
Communicate · Principle 05

Don't perfect the prompt. Frame the conversation.

The session is the unit, not the message.

22 · Frame
Communicate · Principle 06

Great examples beat great instructions.

Show what good looks like. Show what bad looks like.

23 · Examples
Communicate · Principle 07

Ask AI to ask you questions.

Socratic beats one-shot. Every single time.

24 · Questions
Evaluate · Principle 10

AI wants to give you a quick win. Don't take it.

Push back. Ask what it's holding back.

25 · Don't settle
Laptops open →

Lab 02 · Prompt Builder

talkingtomachines.xyz/tools/prompt-builder
26 · Hands-on
Homework · 1 of 4

Prompt the back office.

  • Pick two chores from your own list: spreadsheets, minutes, invoices.
  • Build a reusable prompt for each in the prompt builder.
  • Save both. They become skill material in Session 2.
27 · Homework
Principle 08 · Recursion

Ask AI for help with AI.

Use the machines on each other.

28 · Recursive
Homework · 2 of 4

Ask AI to write your prompts.

  • Learn how different media need different prompts: an image, an Instagram post, a song, a research question, a slide show.
  • Ask AI to help you.
cheat codes: talkingtomachines.xyz/field-guide
29 · Homework
Evaluate · Principle 09

Iteration is everything.

Three to five passes. Every time.

30 · Iteration
Homework · 3 of 4

One image. Five versions.

  • Pick one product shot and generate it.
  • Then regenerate four more times, changing one thing in the prompt each pass.
  • Keep the best version, and note what you changed to get there.
31 · Homework
Evaluate · Principle 11

Never trust one machine.

Same task. Three models. Compare.

32 · Cross-check
Homework · 4 of 4

One cross-check.

  • Same task, three machines: Claude, ChatGPT, Gemini.
  • Note where they agree, and where they diverge.
  • Let a colleague judge the winner, blind.
33 · Homework
Evaluate · Principle 12

Always humanize the output.

Nothing ships raw. Deslop every time.

34 · Deslop
Laptops open →

Lab 03 · De-slop Lab

talkingtomachines.xyz/tools/deslop
35 · Hands-on
Principle 14 · The long game

Your sessions become your SOP.

The conversations are the corpus. Guard them.

36 · SOP

The fifteen.

All of them, one page →
01Learn for the AI of tomorrow.
02Not technical. It's communication.
03Choose → Communicate → Evaluate.
04Pick the right AI for the task.
05Frame the conversation, not the prompt.
06Great examples beat great instructions.
07Ask AI to ask you questions.
08Ask AI for help with AI.
09Iteration is everything. 3 to 5 passes.
10Don't take the quick win.
11Never trust one machine. Cross-check.
12Always humanize. Deslop everything.
13Conviction. Art. Sense. Judgment.
14Your sessions become your SOP.
15Every task = Know · Do · Show.
37 · The fifteen
Homework · all four

Before Session 2.

  • Two saved back-office prompts.
  • Five prompts written by AI, one per medium.
  • One image, five versions.
  • One cross-check across three machines.
  • And as a team: which AI do we standardise on? Bring the pick.
38 · Homework
Session 2 →

Agents, skills and a plan.

Next time, everyone in this room builds real software.

39 · Fin