1 · Maturity diagnostic
10 minOpen the project diagnostic and answer the eight questions thinking about the AI you use most today.
Done when: you wrote down your score, the place to start, and which kit the diagnostic recommended.
Lesson 1 · 14 min
The seed and the garden: why AI is cultivated
Nobody writes, line by line, a model's ability to code or plan. The labs create the conditions and the abilities emerge. You do the same on a different scale: the model arrives ready-made, and what you cultivate is the environment around it.
Start the lesson
In this lesson you will
- 1Why the model is the seed and doesn't change with use
- 2What you cultivate: the envelope around the model
- 3The cycle that turns errors into a better agent
- 4Where the garden metaphor breaks down
Goal: separate what the lab cultivates (the model) from what you cultivate (the agent) and explain the cycle that makes the agent improve.
01The cultivation cycle
Each turn is a harvest cycle. Click each step.
Scroll the page: each step below lights up its part of the diagram.
- 1/7
Process
Everything starts with a real process: qualifying leads, organizing the week, scheduling appointments. Without a defined process, there's nothing to evaluate later.
- 2/7
Agent
The agent is the model (fixed, from the lab) plus the environment you wrote: role, context, tools, rules, examples, and memory.
- 3/7
Execution
The agent runs the task with the environment as it stands at that moment. This is where most people stop: they use it, like it or don't, and start over.
- 4/7
Result
What was delivered. It can look good and be wrong: fluent text isn't correct text.
- 5/7
Evaluation
Someone measures it: right, partial, or wrong, and why. Without evaluation, the agent repeats the same error with the same confidence.
- 6/7
Feedback
The evaluation becomes a concrete change in the environment: a new rule, an example, a limit. In the file, not in the conversation.
- 7/7
What changes from one harvest cycle to the next
It's not the model. It's what you noted about where it went wrong and what you adjusted in the environment. The first four steps make the agent work; evaluation and feedback make it improve.
Before, we programmed the behavior. Now, we program the process that produces the behavior.
02The two levels of cultivation
The point almost everyone mixes up.
1. Level 1
Training, weights, and scale happen at Anthropic, OpenAI, Google, DeepSeek. It costs billions and you don't take part.
The result is the seed: abilities that emerged from training, some expected, others that surprised even the people who trained it.
2. Level 2
The model arrives ready-made, with its weights frozen. What you cultivate is everything around it: role, context, tools, rules, examples, memory, evaluation, and feedback.
That envelope is what turns the same model into either a confused intern or a reliable professional.
Consequence
In use, the model doesn't learn on its own. When you say "my AI got better," what improved was the garden. And the garden is yours.
3. Software × cultivation
With traditional software, you write the exact rules and get exactly what was specified; an error is a bug on a single line.
With cultivated AI, you write the conditions and watch what emerges, and an error calls for a different environment: more context, another example, a new rule.
4. Your role
You don't need to be better than the AI at the task. You need to know how to create the environment where it produces the right result.
The work shifts from doing to defining the role, giving context, setting limits, evaluating, and giving feedback. It's people management, applied to systems.
03Video: why your AI repeats the same mistake
Ten minutes with Nei on the whole concept, with chapters and captions.
Cultivated AI: why your AI keeps making the same mistake · 9 min 54 sVideo page with script →
04Where the metaphor breaks down
Read this before you get too excited.
1. It doesn't grow on its own
A plant grows on its own; an agent doesn't. Without evaluation and feedback, it repeats the same mistake forever.
2. It invents fruit
A plant doesn't invent fruit. An agent can produce a result that looks good and is wrong. That's why evaluation exists.
3. It takes work
Bad context, a bad example, and a vague rule produce a bad agent. The model is the same as your competitor's; the garden is not.
4. It's not a conversation
"Giving feedback" here means editing the environment, not chatting until it "understands."
05Two templates to start
Copy them and use them in the practice below.
Test question for your AI
What do you do differently today compared to last month, and why? If you can't answer with examples, say so clearly.
Note: the correction that keeps repeating
# Corrections I keep repeating | what I explain again | how many times | where it should live (context / rule / example) | |---|---|---| | e.g.: answer in short bullet points | 4 | rule | | | | |
06Practice
Start by measuring where you are.
2 · The correction that keeps repeating
10 minThink of something you've already explained to the AI three or more times (a format, a tone, a piece of data). Write it in a note: what you keep repeating and which of the eight elements it should live in (context, rule, or example).
Done when: the note has the repeated correction and the element where it will live.
07Quick quiz
4 questions. The explanation appears after you try.
Question 1 of 4
After a month of use, your AI makes fewer mistakes. What changed?
Correct
In use, the weights stay frozen. What learns is the system: the files you write and review. The common mistake is thinking the model "learned from you."Not this one
In use, the weights stay frozen. What learns is the system: the files you write and review. The common mistake is thinking the model "learned from you."Review: The two levels of cultivationQuestion 2 of 4
Which step of the cycle do most people skip?
Correct
Using it and starting over is the default. Without measuring the result and feeding the correction back into the environment, the cycle doesn't close and the agent doesn't improve.Not this one
Using it and starting over is the default. Without measuring the result and feeding the correction back into the environment, the cycle doesn't close and the agent doesn't improve.Review: The cultivation cycleQuestion 3 of 4
What does "giving feedback" mean in cultivation?
Correct
Feedback is a concrete change to the file. Repeating the explanation in the chat disappears in the next session.Not this one
Feedback is a concrete change to the file. Repeating the explanation in the chat disappears in the next session.Review: The two levels of cultivationQuestion 4 of 4
Two companies use the same model. Why are the results so different?skill
Correct
The model is available to everyone. The difference lies in the context, rules, examples, and evaluation that each company has built.Not this one
The model is available to everyone. The difference lies in the context, rules, examples, and evaluation that each company has built.Review: The cultivation cycleQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.Choose an option. The explanation appears after your attempt.
08Summary
Seed and gardenThe lab cultivates the model; you cultivate the agent.
The cycleProcess, agent, execution, result, evaluation, feedback, better agent.
In the fileFeedback is editing the environment, not repeating it in the chat.