Catalog of Concepts and Learning Objectives

How to use this catalog

This catalog lists, for each module, the concepts it teaches, the learning objectives a student should meet, and what fluency looks like: concrete things a student who has completed the module should be able to do without help. Use it to plan, to check understanding before moving on, and to decide whether a student has completed the curriculum at a satisfactory level.

The concepts appear as definitions so that teachers and students share the same vocabulary. The objectives appear as things a student does. The fluency examples are the kind of task a student should be able to sit down and do with an AI tool.

The progression

The curriculum moves through five foundational concepts. At each point, a student should be able to do everything before it.

After module	The student can
1. Programming AI with Language	Write a prompt that is a program, supply inputs to it, and update it.
2. Giving AI Context	Identify what types of information an LLM can and cannot know from its training data, supply new information to the LLM in a prompt, understand what a hallucination is, and write rules for how the LLM should handle ambiguity, missing information, or contradictions.
3. Multimodal AI Context	Use a picture as context and perform basic visual operations on image-based context.
4. Instructing AI by Example	Understand example-based prompts, in-context learning, chain of thought, and style replication by example.
5. Programming AI with Personas	Describe what a persona is, how a persona changes the behavior of an LLM, and how to design and use personas to guide the LLM's work.
6. Organizing & Processing Information with AI	Extract, transform, and organize information in a variety of forms in order to synthesize new outputs such as outlines, tables, and analyses.

A student who can do all six has met the curriculum's outcome: building and solving sophisticated problems with large language models through prompt engineering.

Module 1: Programming AI with Language

Concepts

Prompt. The message a person sends to an LLM. A prompt can ask a question or tell the LLM what to do.
Prompt-based program. A set of instructions, written in ordinary language, that tells the LLM what to do with the information it receives.
Instruction for a future point in the conversation. A prompt written with language such as whenever I give you…, from now on…, or after I provide…, so that it applies to inputs that arrive later in the conversation.
Input. A message that provides the information the program acts on, as opposed to a message that changes the program.
Conditional rules. Instructions for what the LLM should do in a particular situation: after the player guesses, when the topic is missing, when the input is not a number.
State. Information the program must remember from one turn to the next, such as a score or a round number.
Output format. A description of exactly how the LLM's responses must look, so that they are predictable: numbered statements, a fixed set of labeled lines, only the result.
Updating a prompt-based program conversationally. Adding, changing, or removing a rule in a later message, without restating the whole program.

Learning objectives

Write an instruction, in a conversation with an LLM, that tells it what to do at a future point in the conversation. Then follow up in the conversation and provide the input it needs to perform the computation.
Express the logic of a program in language: the steps in order, the rules for what the LLM should do in particular situations, what it must remember between turns, and the format its responses must follow.
Update a prompt-based program conversationally by sending a new or changed rule, and confirm the changed behavior by providing the same input again.
Test a prompt-based program by running it, noticing where its behavior differs from what the student intended, and revising the prompt to close the gap.
Apply the same approach to different kinds of input, such as numbers, passages of text, observations, topics, or the messages a player sends in a game.

Fluency looks like

Building a game such as Fact or Fiction from scratch: a starting prompt with the steps, a topic input, a guess input, a rule for each, a score that persists across rounds, a rule for continuing, and a fixed format for every round and result.
Writing a program that turns any passage into flashcards, sending three passages one at a time, and then adding a difficulty rating with a single follow-up message.
Designing an original game, running it, identifying what is confusing or missing, and improving the prompt across three versions, with each change described in a sentence or two.
Taking the same approach to something that is not a game: a program that organizes observations into a table, or one that turns a dish name into a priced shopping list.

Module 2: Giving AI Context

Concepts

Context. All the information in the conversation that the LLM can use to do the task: the input, the instructions, the purpose, the output format, and the rules.
LLM training. An LLM learns by reading text and predicting the next word, over an enormous amount of text. From this it learns patterns in language and the concepts those patterns express.
Next-word prediction and pattern recognition. When an LLM responds, it completes patterns it learned in training, using whatever is in the conversation. "Mary had a little ___" completes one way for everyone. "My name is ___" completes with whatever name the conversation supplied.
An LLM's two sources of information. What it learned in training, and what the conversation contains. Anything that happened after training, anything nobody wrote down, and anything private to the student stays unknown to it unless the conversation supplies it.
Hallucination. A plausible but unsupported answer that the LLM produces when it lacks the information to answer and fills the gap with a guess.
Rules for missing information and unexpected situations. Instructions in the prompt for what the LLM should do when information is missing, an input is not what the program expected, an instruction is ambiguous, or two rules contradict each other.

Learning objectives

Explain next-word prediction in the LLM training process, and how it leads the LLM to recognize and complete patterns using the context in the conversation.
Identify what types of information an LLM can and cannot know from its training data, and supply the information it cannot know in a prompt.
Understand what a hallucination is, recognize one in a response, and explain what context was missing.
Write a prompt that gives the LLM a job, the input to work on, what matters about the result, and the output format, so that the result fits the student's actual situation instead of a typical one.
Write rules for how the LLM should handle missing information, invalid or off-task input, ambiguity, and contradictions, so that it asks or reports instead of guessing.

Fluency looks like

Building a study planner that takes the student's own test date, topics, and available time, specifies the plan's format, and asks for anything it needs instead of inventing it.
Building a quiz game from the student's own notes, with rules for the question types, what happens on an invalid answer, what happens on off-task input, and how the LLM chooses the next question.
Taking a request that the LLM answered wrongly, identifying the context it never had, and writing the message that should have come before it.
Doing the same for any source of personal information: a syllabus, a practice schedule, a set of class notes, a family calendar.

Module 3: Multimodal AI Context

Concepts

Multimodal context. Context that comes in more than one form. Text is one form. A picture is another.
Image-based context. A photograph or drawing supplied as context alongside a written instruction. It carries details nobody would have typed and shows the LLM the student's physical situation.
Visual operations. The basic things a person can direct the LLM to do with image-based context. This curriculum teaches five.
Extract. Pull the wanted pieces of information out of a picture and leave the rest, such as an inventory from a fridge or a table from a pile of receipts.
Interpret. Read a picture against a question or goal, such as a menu against dietary goals, a control panel against a task, or a draft against a reviewer's eye.
Transform. Turn what is in a picture into something new, such as ingredients into a recipe or an object into build instructions.
Simulate. Treat a drawing as a working thing and run it, such as a maze or a machine that becomes a game.
Extend. Continue a picture that is unfinished, such as a half-filled schedule.

Learning objectives

Provide a photograph or drawing as context, together with a written instruction that says what to do with what the LLM can see.
Perform each of the five visual operations on image-based context by writing the prompt for it.
Choose the operation that fits a goal: an inventory calls for extract, a decision against a goal calls for interpret, a recipe from ingredients calls for transform.
Write rules for what the LLM should do about anything in the picture it cannot read or see, and check any assumption it states about what is not visible.

Fluency looks like

Photographing a pantry, shelf, or pile of receipts and producing an inventory or an expense table.
Photographing a schedule, a menu, or a control panel and producing a plan, a choice that fits a stated goal, or step-by-step instructions.
Drawing a game or a machine on paper, photographing it, and directing the LLM to run it so the class can play.
Taking two pictures of the student's own and climbing the whole ladder: extract, interpret, review, transform, simulate, extend. Then doing the same with any object, document, or drawing.

Module 4: Instructing AI by Example

Concepts

Instruction versus example. Two ways to teach: state the rule, or show correctly completed work and let the LLM infer the rule.
In-context learning. The LLM learns how to perform a task from the examples in the prompt, not from any change to the model. Also called example-based prompting.
Few-shot prompting. A prompt that includes a few examples of the task before asking for a new case. A prompt with no examples is zero-shot.
Classification. Putting something into a category by giving it a label. A handful of labeled examples can teach a classifier.
Boundary. The line between what belongs in a category and what does not. One example on the edge teaches it.
Procedure by example. Showing the steps of a process, or the form of a result, so the LLM performs the same process on a new case.
Chain of thought. The sequence of reasoning steps the LLM writes down on the way to an answer. Because each step becomes context for the next, a correct chain makes a correct answer far more likely. "Think step by step" is the simplest way to ask for one.
Worked example. Showing the desired chain of thought, step by step, so that the LLM reasons that way rather than its own.
Style replication by example. Teaching a voice or style of writing from samples, which rules cannot capture.
Refining by pointing. Improving a prompt by quoting sentences from its output as preferred and not-preferred examples.

Learning objectives

Write a few-shot prompt: input and output pairs in a consistent format, followed by a new input for the LLM to complete, with no rule stated.
Build a classifier for something the student sorts in their own life, find how few examples it takes, and add one example that teaches a boundary the others left out.
Teach a procedure or a format by example, including one the LLM has never seen, such as a made-up language.
Ask for a chain of thought with think step by step, and then design the chain with a worked example so the LLM reasons the student's way.
Replicate a style from samples of writing, and refine any result by quoting what to keep and what to drop.

Fluency looks like

Building an example library: a classifier, a procedure, and a refinement, each taught with examples alone, and a short declaration of what each one does.
Diagnosing why a shopping-list extractor included a person, and adding the one example that fixes it.
Writing a worked example that shows how to check whether an article supports its headline, and applying it to a new headline.
Taking a flyer or an essay the LLM wrote, marking the sentences to keep and the sentences to drop, and getting a better version. Then doing the same for any task where the rule is hard to write.

Module 5: Programming AI with Personas

Concepts

Persona. A description of the person, thing, system, or role the LLM should act as when it does the work.
Persona prompt pattern. Act as [description]. A short description carries a large pattern of behavior the LLM learned in training, so nobody has to spell the behavior out.
Role-based persona. A job or perspective, such as a scientist or a nutritionist. A job description is a good starting point.
Personality-based persona. A trait that changes how the LLM approaches the work, such as a skeptic or a supporter.
Example-driven persona. A persona taught from samples of how someone talked or behaved.
Persona as operation, structure, or whole. A persona can do something to an idea, decide the shape of an answer, or become the entire thing, including its rules, controls, memory, and screen.
State and named fields. What a persona that acts as a system must remember from one turn to the next, made visible by writing the prompt as named fields for its inputs and responses.
Language as program and interface. With generative AI, words can describe the thing, state its rules, serve as its controls, and draw its display.

Learning objectives

Describe what a persona is and how giving the LLM a persona changes what it notices, how it works, and what it produces.
Design role-based, personality-based, and example-driven personas for the same problem, make a persona more specific, and predict how each guides the LLM's work differently.
Choose the persona that should do a piece of work, and state the reason.
Build a persona that acts as a system: describe the thing, its rules, its inputs, its state, and the fields of its responses. Change its rules while it is running, break it on purpose, and repair it.

Fluency looks like

Giving the same science project to a skeptic and a supporter and explaining what each noticed.
Making the LLM act as an animal, then a stranger with a specific background, then a system such as a vending machine with named fields for credit, action, and next input.
Building something that is not a person, such as an arcade game or a pinball machine, playing it, changing the physics in a sentence, and describing what held and what broke.
Using the customers' own messages to make the LLM act as those customers and react to a change the student is considering. Then doing the same for any reviewer, audience, or system the student needs to hear from.

Module 6: Organizing & Processing Information with AI

Concepts

Extract. Pull specific pieces of information out of something larger and leave the rest.
Transform. Change information from one form into another: one unit system, a plain name, a table row, a short bullet.
Organize. Arrange information in an order or in groups. The way something is organized is a choice that decides what the reader sees.
Synthesis. Combining extract, transform, and organize to produce a new output from the information: an outline, a table, a plan, or an analysis.
Outline. A structure with headings and bullets that extracts, transforms, and organizes at once.
Decomposition. Any bullet is a piece of information in its own right and can break down further.
Where the LLM filled the gap. When the LLM adds something that was not in the input, it came from context the student provided, from training, or from invention. Each can be useful. Each needs noticing.

Learning objectives

Write an extract instruction that names the kind of pieces wanted, and apply it to notes, receipts, messages, transcripts, or articles.
Write a transform instruction that names the form wanted, and check what the transform lost.
Write an organize instruction that names the order or the groups, and try a second organization when the first does not answer the question.
Combine all three in one instruction to synthesize a new output such as an outline, a table, a plan, or an analysis, and expand any part of it further.
Identify where the LLM filled a gap from context, from training, or by invention, and add instructions that make it mark anything not in the input.

Fluency looks like

Turning a group chat into a plan: extract the people and what each is bringing, transform the times, organize by person and then by time, and then do it all in one instruction.
Turning a pile of receipts into an expense table with totals by category, then reorganizing by month, then asking what the totals reveal.
Turning a voice transcript of a planning conversation into an outline of decisions, tasks, and open questions, and expanding one bullet into detail.
Taking one bullet of an outline and marking which added details came from the notes, from general knowledge, or from nowhere. Then applying the three instructions to any information that arrives in a hard-to-use form.

Satisfactory completion

A student has completed the curriculum at a satisfactory level when they have completed all six modules with their teacher, have turned in all six assignments with each part completed as the step asked, and can do the fluency tasks above for each module. A student who meets this standard has earned recognition, such as a Certificate of AI Foundations from the school.