field notes

Field notes

42 field notes in 11 clusters, plus a start-here guide. Each is one concept, one true story from our own systems and a two-minute experiment. The 9 free foundations are readable now. The rest unlock in the program.

Join the cohort 1 waitlist About the program

The path

The same notes in the order the program uses them.

Week 0

0.1How should I use these notes?
Forty-two short notes, two ways through them, and one trick that makes them yours.
4 min · Free
1.1What does a forward-deployed engineer actually do?
The job title sounds like software, and the code turns out to be the smallest slice of it.
10 min · Free
1.2Why does the demo always work and production always surprise you?
A demo runs on the data you chose. Production runs on all of it, and all of it includes the rows you never looked at.
10 min · Free
3.1What is a model actually doing when it answers you?
It has exactly one trick, repeated very fast. Once you see the trick, you'll understand both why it's so good and why it sometimes makes things up with a straight face.
10 min · Free
3.2What are tokens, and why does anyone count them?
Every AI system has a meter running in both directions. Most people only discover this when the bill arrives.
10 min · Free
3.3What is a context window, and what happens when it fills up?
A model can only consider what's in front of it right now, and the space in front of it has a hard edge. What falls past that edge simply doesn't exist for it.
10 min · Free
3.4Why does the same question get different answers?
A system that gave the right answer once has told you almost nothing. Here is why, and what we do about it.
10 min · Free

Warm-up

Project 1

1.3Who pays when the AI is wrong?
Mistakes are not equal in price. Find the one that costs the most and build the system around stopping that one.
10 min · In the program
2.1How do you find the problem actually worth automating?
Three numbers decide whether a task deserves automation, and sometimes the right answer is to walk away.
10 min · In the program
2.2What does a good requirement one-pager look like?
One page, six headings, and a single hard constraint that can reshape the whole system behind it.
10 min · In the program
4.1Why do vague prompts produce confident garbage?
The model will always answer. When you leave gaps in the request, it fills them with the most average guess available, in a perfectly assured voice.
10 min · In the program
4.2What is a structured output, and why do businesses insist on it?
A database can't store a paragraph. Getting a model to answer in fields instead of prose is where prompt work stops being writing and starts being design.
10 min · In the program
6.2Why does every agent need a budget?
A loop that costs a little per turn has no ceiling unless you build one. A budget turns a nasty surprise into a number you chose.
10 min · In the program
7.1Where does an agent keep its state?
The model remembers nothing between calls. Everything it appears to know was put in front of it, again, by your program.
10 min · In the program
7.2What should the model not decide?
Every question you hand to a model costs money, takes time and can come back wrong. A surprising number of them never needed a model.
10 min · In the program

Project 2

4.3How do examples teach a model what you mean?
Some instructions can't be explained but can be shown. A handful of well-chosen examples often does more than a page of rules.
10 min · In the program
4.4How do you know your new prompt is better and not just different?
Every edit to a prompt feels like an improvement. A fixed set of hand-checked answers is the only thing that tells you when it isn't.
10 min · In the program
6.4What happens when you run it twice?
Everything you build will be run twice by accident. The only question is whether the second run quietly doubles your data.
10 min · In the program
6.5Why does your agent need a harness of its own?
A harness is a program whose whole job is to distrust your system. Building one is the most employable skill in this course.
10 min · In the program
7.4What is a skill, and how does an agent learn your way of working?
A skill is a file of rules and real examples that an agent loads when the task matches. It is how taste gets passed on without retraining anything.
10 min · In the program
8.2Why is real-world data always messier than the demo data?
A field can be present, well-formed and wrong. Data lies in layers, and each layer needs its own check.
10 min · In the program
9.2How should a system fail?
The most expensive output a pipeline can produce is a green checkmark that isn't true.
10 min · In the program
10.1Why doesn't the model learn from its mistakes?
The model you use today is the same model tomorrow, however many times it gets something wrong. The system around it is where learning happens, if you build it.
10 min · In the program
10.2What is a decision trace, and why write down the why?
Logs tell you what your system did. A decision trace tells the next person why it was built that way, and that is the part nobody can reconstruct later.
10 min · In the program

Project 3

5.1Why can't you just paste all the company's documents into the prompt?
It sounds like the obvious design, and it fails three different ways before the documents even fit.
10 min · In the program
5.2What is retrieval, and what is an embedding?
Search that matches words fails the moment someone phrases things differently. Search that matches meaning has to turn meaning into numbers first.
10 min · In the program
5.3Why must a knowledge agent sometimes refuse to answer?
A system that always has an answer will eventually give you a wrong one in a perfectly confident voice.
10 min · In the program
5.4How do you measure a system that answers questions?
One accuracy number hides three different ways to be wrong. Grade them separately and you'll know what to fix.
10 min · In the program
7.3Why must data never give orders?
Any text your system reads can contain a sentence addressed to your model. Safety comes from how you build, not from asking nicely.
10 min · In the program

Project 4

2.3Why do stakeholders say "automate everything" and mean "don't touch anything"?
People ask for full automation and then get nervous at the first sign of it. Trust arrives in slices.
10 min · In the program
6.3What is an approval gate, and where do you put one?
You can't ask a human to approve every step, and you can't let an agent do everything alone. The trick is knowing which steps cannot be taken back.
10 min · In the program
7.5How does an agent remember yesterday?
Agent memory is files it writes and reads back. That makes it simple to build, and it makes stale memory a liability.
10 min · In the program
9.1What must the model never see?
Anything you put in a prompt has left the building. The safest data to protect is the data you never sent.
10 min · In the program
9.3Who approved that? Audit trails and the paper they leave
When an AI system gets something wrong, the first question is which rows it touched. A system that can't answer that has to redo everything.
10 min · In the program

Capstone

7.6When does one agent become several?
Splitting work across agents adds a cost, coordination. Pay it only when the split earns it.
10 min · In the program
8.3What breaks when two systems quietly disagree?
Both systems pass their tests and the combination is still wrong. The bug lives in the assumption between them.
10 min · In the program
8.4How does your agent plug into tools it has never seen?
APIs let programs talk, but every connection used to be custom-built. A shared protocol changed what one small team can ship.
10 min · In the program
10.3Should the agent be allowed to improve itself?
An agent can read its own track record and suggest what to change. Whether those suggestions take effect should never be up to the agent.
10 min · In the program
11.2How do you present numbers people can check?
The strongest number you can show is the one the reader could reproduce. Sizing your claim to your evidence is a skill.
10 min · In the program
11.3How do you say "it didn't work"?
A clean account of a failure is worth more than a clean record. Here is a real one, written the way we write them.
10 min · In the program

Every project

11.1How do you explain your system to someone who will never read the code?
The person who decides whether your system gets used will read one page. Here is the page.
10 min · In the program

By cluster

0 Start here

0.1How should I use these notes?
Forty-two short notes, two ways through them, and one trick that makes them yours.
4 min · Free

1 AI deployment mindset and FDE responsibilities

1.1What does a forward-deployed engineer actually do?
The job title sounds like software, and the code turns out to be the smallest slice of it.
10 min · Free
1.2Why does the demo always work and production always surprise you?
A demo runs on the data you chose. Production runs on all of it, and all of it includes the rows you never looked at.
10 min · Free
1.3Who pays when the AI is wrong?
Mistakes are not equal in price. Find the one that costs the most and build the system around stopping that one.
10 min · In the program

2 Business problem discovery and requirement mapping

2.1How do you find the problem actually worth automating?
Three numbers decide whether a task deserves automation, and sometimes the right answer is to walk away.
10 min · In the program
2.2What does a good requirement one-pager look like?
One page, six headings, and a single hard constraint that can reshape the whole system behind it.
10 min · In the program
2.3Why do stakeholders say "automate everything" and mean "don't touch anything"?
People ask for full automation and then get nervous at the first sign of it. Trust arrives in slices.
10 min · In the program

3 GenAI and LLM foundations

4 Prompt engineering for business use cases

4.1Why do vague prompts produce confident garbage?
The model will always answer. When you leave gaps in the request, it fills them with the most average guess available, in a perfectly assured voice.
10 min · In the program
4.2What is a structured output, and why do businesses insist on it?
A database can't store a paragraph. Getting a model to answer in fields instead of prose is where prompt work stops being writing and starts being design.
10 min · In the program
4.3How do examples teach a model what you mean?
Some instructions can't be explained but can be shown. A handful of well-chosen examples often does more than a page of rules.
10 min · In the program
4.4How do you know your new prompt is better and not just different?
Every edit to a prompt feels like an improvement. A fixed set of hand-checked answers is the only thing that tells you when it isn't.
10 min · In the program

5 RAG systems and enterprise knowledge workflows

5.1Why can't you just paste all the company's documents into the prompt?
It sounds like the obvious design, and it fails three different ways before the documents even fit.
10 min · In the program
5.2What is retrieval, and what is an embedding?
Search that matches words fails the moment someone phrases things differently. Search that matches meaning has to turn meaning into numbers first.
10 min · In the program
5.3Why must a knowledge agent sometimes refuse to answer?
A system that always has an answer will eventually give you a wrong one in a perfectly confident voice.
10 min · In the program
5.4How do you measure a system that answers questions?
One accuracy number hides three different ways to be wrong. Grade them separately and you'll know what to fix.
10 min · In the program

6 Agentic AI workflows and automation design

6.1What turns a chatbot into an agent?
The jump from answering to acting is small in code and enormous in consequences.
10 min · Free
6.2Why does every agent need a budget?
A loop that costs a little per turn has no ceiling unless you build one. A budget turns a nasty surprise into a number you chose.
10 min · In the program
6.3What is an approval gate, and where do you put one?
You can't ask a human to approve every step, and you can't let an agent do everything alone. The trick is knowing which steps cannot be taken back.
10 min · In the program
6.4What happens when you run it twice?
Everything you build will be run twice by accident. The only question is whether the second run quietly doubles your data.
10 min · In the program
6.5Why does your agent need a harness of its own?
A harness is a program whose whole job is to distrust your system. Building one is the most employable skill in this course.
10 min · In the program

7 The anatomy of an agent

7.1Where does an agent keep its state?
The model remembers nothing between calls. Everything it appears to know was put in front of it, again, by your program.
10 min · In the program
7.2What should the model not decide?
Every question you hand to a model costs money, takes time and can come back wrong. A surprising number of them never needed a model.
10 min · In the program
7.3Why must data never give orders?
Any text your system reads can contain a sentence addressed to your model. Safety comes from how you build, not from asking nicely.
10 min · In the program
7.4What is a skill, and how does an agent learn your way of working?
A skill is a file of rules and real examples that an agent loads when the task matches. It is how taste gets passed on without retraining anything.
10 min · In the program
7.5How does an agent remember yesterday?
Agent memory is files it writes and reads back. That makes it simple to build, and it makes stale memory a liability.
10 min · In the program
7.6When does one agent become several?
Splitting work across agents adds a cost, coordination. Pay it only when the split earns it.
10 min · In the program

8 APIs, data pipelines and system integration

8.1What is an API, really?
Every useful AI system spends most of its life asking other programs questions. The rules for asking are older and simpler than you think.
10 min · Free
8.2Why is real-world data always messier than the demo data?
A field can be present, well-formed and wrong. Data lies in layers, and each layer needs its own check.
10 min · In the program
8.3What breaks when two systems quietly disagree?
Both systems pass their tests and the combination is still wrong. The bug lives in the assumption between them.
10 min · In the program
8.4How does your agent plug into tools it has never seen?
APIs let programs talk, but every connection used to be custom-built. A shared protocol changed what one small team can ship.
10 min · In the program

9 AI governance, privacy, risk and guardrails

9.1What must the model never see?
Anything you put in a prompt has left the building. The safest data to protect is the data you never sent.
10 min · In the program
9.2How should a system fail?
The most expensive output a pipeline can produce is a green checkmark that isn't true.
10 min · In the program
9.3Who approved that? Audit trails and the paper they leave
When an AI system gets something wrong, the first question is which rows it touched. A system that can't answer that has to redo everything.
10 min · In the program

10 Systems that improve themselves

10.1Why doesn't the model learn from its mistakes?
The model you use today is the same model tomorrow, however many times it gets something wrong. The system around it is where learning happens, if you build it.
10 min · In the program
10.2What is a decision trace, and why write down the why?
Logs tell you what your system did. A decision trace tells the next person why it was built that way, and that is the part nobody can reconstruct later.
10 min · In the program
10.3Should the agent be allowed to improve itself?
An agent can read its own track record and suggest what to change. Whether those suggestions take effect should never be up to the agent.
10 min · In the program

11 Stakeholder communication and solution presentation

11.1How do you explain your system to someone who will never read the code?
The person who decides whether your system gets used will read one page. Here is the page.
10 min · In the program
11.2How do you present numbers people can check?
The strongest number you can show is the one the reader could reproduce. Sizing your claim to your evidence is a skill.
10 min · In the program
11.3How do you say "it didn't work"?
A clean account of a failure is worth more than a clean record. Here is a real one, written the way we write them.
10 min · In the program