Cultivated AI · Part One · facilitator plan
Eight hours over four two-hour sessions. Hands on the keyboard throughout, on each leader's own real work. This is the internal run-sheet: timings, the exact activity at each beat, and what each session leaves behind.
How LLMs actually work, so context, cost, and model choice follow from understanding.
The full discipline, practised on each leader's real tasks.
Turn your best repeated work into a reusable skill. First taste of agents.
From one desk to the team to the company, and a written plan to commit to.
Build true intuition for what a model is. Every later decision (how much context, which model, what it costs) should fall out of understanding, not folklore.
Each leader leaves able to say, in their own words, what an LLM does and why it behaves the way it does, and with a cost-and-model cheat sheet for their own top tasks.
Open: this is a people change, not a software install Discuss
Set the room. Leaders, not engineers. The honest frame: AI makes your judgment more valuable, and the parts of the job you actually like come back. Name the fear out loud so it stops running the rollout.
The brain, simplified Teach
A neuron takes signals, weighs them, fires. Stack billions and you get pattern recognition. Walk from the biological neuron to the artificial one on the whiteboard. New visual: neuron to perceptron.
From neural nets to LLMs Teach
What "predict the next token" really means. Training versus inference. The key reframe for the whole workshop: this is a communication tool, not a database to query.
Context is everything Lab
Run Context Window Viz (Ch 3): watch messages fill the window and the oldest fall away. Then Forgetting Experiment (Ch 3): the same chat with and without a summary, so they feel why context management matters.
Break
Short reset.
Tokens and token economics Lab
What a token is, why long chats cost more and drift, and the survival strategies. New widget: Token Cost Calculator, price the same Baker's Dozen task (say a city-sales digest) across models and see the spread.
Choosing the right model Lab
The big three (ChatGPT, Claude, Gemini), when each wins, and Principle 4: when in doubt, ask AI which AI to use. Run the same prompt three ways (the Big Three shootout). New widget: model-selection decision tree.
First principle on the keyboard Lab
Principle 5, frame the conversation. Each leader runs one real task the right way: a session, not a one-line prompt. Sets up Session 2.
A working mental model of how LLMs behave, plus a filled-in cost-and-model cheat sheet for their three most frequent tasks.
From "using AI" to a repeatable discipline. Every principle is practised on the leader's own real work, not a toy example.
Each leader applies the full fifteen-principle loop to three or more of their real tasks, and leaves with a personal cheat sheet they will actually reach for.
Recap and the core loop Teach
Principles 1 to 3: build around approach not tools; it's a communication problem, not a technical one; Choose → Communicate → Evaluate, on repeat. The session is the unit, not the message.
Frame, and show what good looks like Lab
Principle 5 (frame the conversation) and 6 (great examples beat great instructions). Run Prompt Makeover (Ch 1): toggle the building blocks, watch quality climb. Then each leader frames a real brief (a Blinkit listing, a vendor email) with a good and a bad example attached.
Make the machine interrogate you Lab
Principle 7 (ask AI to ask you) and 8 (ask AI for help with AI). Run Flip the Script (Ch 2): the AI asks clarifying questions before it answers. Compare the Socratic result to the one-shot answer.
Break
Short reset.
Don't settle, iterate Lab
Principle 9 (iteration is everything, three to five passes) and 10 (don't take the quick win, ask what it's holding back). Each leader takes one task through five passes and watches it get genuinely better.
Trust nothing once Lab
Principle 11 (cross-check across three models) and 12 (humanize, deslop, nothing ships raw). Run Sycophancy Test (Ch 9): watch the model cave to a leading question, then ask it neutrally. Deslop one real output together.
The human edge, and a mental model Lab
Principle 13 (conviction, taste, judgment come from you), 14 (your sessions become your SOP), 15 (every task is Know, Do, or Show). Run Taste Test (Ch 10): rank five variants against expert consensus. Then each leader sorts their week into Know / Do / Show.
Close: your principle cheat sheet Discuss
Each leader names the two principles they break most often. That becomes their focus for the next week.
The fifteen principles applied to three or more real tasks, and a Know / Do / Show map of their own working week.
Turn a leader's best repeated work into a reusable skill, then take a first taste of agents that act on their behalf.
Each leader ships one working, voice-tuned skill they will use the next working day, and leaves with a shortlist of daily tasks ready to hand to a browser agent.
What a skill is Teach
Name, trigger, steps, examples. Why a saved skill beats re-prompting from scratch every time, and how it captures what's in your head.
Skills anatomy lab Lab
Deconstruct a real, proven skill into its parts so the shape is obvious before anyone writes one.
Find before you build Lab
Official (Anthropic), community, and your own. Browse, pick one close to a real need, and plan the tweak. Most good skills are adapted, not invented.
Build your first skill, end to end Lab
Each leader writes a SKILL.md for one real Baker's Dozen workflow (a Blinkit listing drafter, a city-sales daily digest, a vendor-negotiation brief), with triggers, steps, and examples, then tunes it to their voice. Includes a 5-minute break mid-build.
Compose skills Lab
Chain two skills into one bigger workflow (digest then draft the action email). The whole becomes more than re-running each by hand.
First taste of agents Demo Lab
The agent loop: Plan → Act → Observe → Refine. Live demo of a Chrome extension that acts in the browser. Then Agent Blueprint (Ch 6): plan a simple daily-task automation step by step.
Close Discuss
Each leader commits to running their new skill once before Session 4.
One working, voice-tuned skill ready for Monday, and a shortlist of repetitive daily tasks to hand to a browser agent.
Move skills from one desk to the team to the company, place each function on the ladder, and leave with a plan to commit to.
Each leader leaves with a written 90-day plan (phases, owners, metrics) and a model for how skills spread between teams instead of staying siloed.
Why good work stays siloed Discuss
Your best person's best workflow usually dies on their laptop. The whole game in Part One's last session is making it spread, so one person's leverage becomes everyone's default.
From desk to team Lab
Package a skill so a teammate can run it cold; the handoff doc is the real interface. New activity: team skill-vote, the table picks one skill and refines it together until anyone can run it.
Across the company Teach
A simple skills library and an ownership model, plus how wins spread (the adoption ladder). Lightweight governance so quality holds as it scales.
Break
Short reset.
Where each team sits Lab
Place each function on the ladder, Quoter → Curator → Author → Orchestrator, and map functions to AI or deliberately not. Honest about where AI does not belong yet.
The 90-day plan Lab
Phase 1 (weeks 1 to 4: foundation and first wins), Phase 2 (weeks 5 to 8: deepen and compound), Phase 3 (weeks 9 to 12: scale, automation, culture). Each leader drafts their own.
Commit Discuss
Success metrics, named owners, a first review date. The concrete path to India's most AI-forward D2C org.
A written 90-day plan with phases, owners, and metrics, and a working model for sharing skills between teams.
Bring real work. The single biggest predictor of stickiness is that every lab runs on the leader's own task, not a sample. The pre-workshop questionnaire exists to collect those tasks in advance, so the room is ready to go.
Pace. Each session keeps teaching short and lab long. If a beat runs over, cut the next teach segment, never the lab.
Carry-over. Sessions are designed to be a week apart, so leaders use what they learned between sittings. Each session opens with a quick "what did you actually run" round.
Room setup. Everyone on a laptop with their account live (see the pre-req list). One large screen for demos. Facilitator runs the new widgets (Token Cost Calculator, model-selection tree, agent demo) from the Talking to Machines site.