AI Prompt Engineering Masterclass

Hey - Welcome In

I'm really glad you're here. In this course, we're going to turn you from someone who "uses AI" into someone who can make an AI do almost anything you need - write like your brand, reason through a messy business problem, review code like a senior engineer, generate a consistent AI video scene, or turn a 40-page report into a one-page answer. None of that is magic. It's a skill called prompt engineering, and like any skill, it's learnable, practiceable, and - once it clicks - genuinely fun.

By the time you finish this course, you'll have your own personal library of reusable prompt templates for the tasks you actually do in your life or job, and you'll never again wonder why the AI "just didn't get it." You'll know exactly what it needed and exactly how to give it.

How This Course Is Built

This course has 98 lectures across 18 modules, split into two parts. Part 1 gives you the core techniques - the skills every other module depends on. Part 2 takes those same skills and applies them to the real content you actually create.

• Part 1 - Core Prompting Foundations (Modules 1–10): Mindset, Zero-Shot, Role Prompting, Few-Shot, Structured Output, Chain-of-Thought, Constraint-Based Prompting, the CRAFT Framework, Task Decomposition, and a Capstone that stacks it all together.
• Part 2 - Applied Prompting Mastery Tracks (Modules 11–18): AI Video, Code Generation, SEO Articles, AI Images, Chatbot System Prompting, Social Media, Email Marketing, and Creative Story & Script Prompting - each one built entirely on what you learn in Part 1.

I've put the full lecture-by-lecture map in a separate one-page roadmap PDF in your resources folder - keep it open in a tab so you always know where you are in the course.

What You Need Before Lecture 1.2

• An AI chat tool open and ready - Claude, ChatGPT, or Gemini all work fine, and the free tier is enough for most of this course.
• A blank document or notebook - this is where you'll paste prompts and keep the ones that work. This becomes your personal prompt library by the end.
• A willingness to actually type along during class, not just watch. This course is built to be done live, not read passively.

How Each Lecture Will Run

Every lecture in this course follows the same rhythm, so you always know what's coming:

• The idea - I explain the technique in plain language, no jargon
• A live example - I show you a weak version and a strong version, side by side
• You try it - a short in-class exercise, usually 3–5 minutes
• Recap + homework - one line to remember, and something small to bring to the next lecture

“This is not a lecture you watch. It's a lecture you do.”

Before Lecture 1.2

The Habit We're Breaking Today

Quick show of hands - mentally, at least. How many of you, when you open an AI chat, type something short, almost like you're searching Google? “best gym marketing ideas.” “dog training tips.” “resume advice.” Totally normal. You've had 15+ years of practice typing short keyword phrases into a search box and getting good results. The problem is: an AI model is not a search engine, and that one habit is the single biggest reason beginners think AI output is “meh.”

A search engine matches your keywords against millions of existing pages and hands you links. It didn't write anything - it just found things. An AI model doesn't search for an existing answer. It reasons, in the moment, based entirely on the instructions you give it. No instructions, no context, no direction - it fills in the gaps with the most generic, average, could-apply-to-anyone answer it can produce.

Let's See It Side by Side

Here are 3 different situations. Same lesson, every time.

Example 1: Gym Marketing

The first one gives you a listicle you've seen a hundred times: post on Instagram, offer referral discounts, run a Google ad. It's not wrong, it's just generic enough to apply to any gym on Earth. The second one gives you ideas that only make sense for a small gym fighting a big-box competitor, because we told it exactly who it's talking to.

Example 2: Dog Training

"Dog training tips" could return generic advice for any dog, any problem. The strong version names the exact behavior, exactly when it does and doesn't happen, and asks for something usable tomorrow, not a general education on dog psychology.

Example 3: Freelance Portfolio

Same pattern a third time. The weak version could apply to anyone building any kind of website. The strong version names the specific person, the specific problem, and the specific outcome, which is exactly why it comes back with tailored advice instead of a generic checklist.

The Test You'll Use for the Rest of This Course

Ask yourself: “Could this exact output work for a totally different person or business with zero changes?” If the answer is yes, your prompt was too thin.

This is the single test I want you to run on every AI output you get from today forward. It will catch 90% of weak prompting on its own, before you even learn a single named technique.

Try It Yourself - Right Now

1. Take the Google search you wrote down for homework.
2. Type it into your AI tool exactly as-is. Read the output.
3. Now rewrite it as a full instruction: add who you are, what the real situation is, and what you actually want to walk away with.
4. Run the second version. Compare. Does the second one pass the “could this apply to anyone” test?

Before Lecture 1.3

A Prompt Isn't One Thing - It's Five

Last lecture we proved that a thin prompt gets a thin answer. Today we're going to open up a strong prompt and look at exactly what it's made of, piece by piece. Once you can see these five parts, you'll never write a prompt “from nothing” again - you'll just fill in five blanks.

• Role - who should the AI act as? What's their expertise?
• Context - what's the real situation or background?
• Task / Action - what exactly do you want it to do?
• Format - how should the answer be shaped or organized?
• Constraints - what should it avoid, limit, or not do?

Keep this list - we're going to come back to these exact five words in Module 8 when we formalize them into a reusable framework called CRAFT. For now, just get comfortable spotting them.

Let's Label a Real Prompt

Here are 3 different fully-labeled prompts, from 3 different fields.

Notice something: none of these are fancier language, longer sentences, or impressive vocabulary. Each one is just five specific decisions the person made before hitting send, and the same five decisions work whether you're reviewing a budget, a workout plan, or a contract. That's the whole secret. Good prompting is a decision checklist, not a writing talent.

Try It Yourself

1. Take the “strong” gym prompt from last lecture.
2. Go line by line and label which part is the Role, which is the Context, which is the Task, and so on.
3. Notice: did it have all five? Which one was missing or weakest? (Most beginner “good” prompts are missing Format and Constraints - that's normal, we'll fix that in Modules 5 and 7.)

Before Lecture 1.4

The Rule Behind Everything We'll Ever Do

Before computers could even talk back to us, engineers had a saying: garbage in, garbage out. It means the quality of what comes out of a system is capped by the quality of what you put into it. No system, no matter how advanced, can turn a vague, low-effort input into a specific, high-effort output. AI models are no exception - if anything, this rule matters more for AI than for almost anything else you've ever used.

Here's the part that surprises people: this isn't about writing more words. It's about writing the right five things we covered last lecture. A short prompt with a clear role, real context, and a specific task will beat a long, rambling prompt with none of those, every single time.

Three Tiers, Three Tasks

Same three tiers, weak, medium, strong, applied to 3 different tasks.

Task 1: Restaurant Recommendation

Task 2: Choosing a Gift

Task 3: Writing a Job Description

Across all three tasks, Tier 2 already feels “good enough” to a lot of beginners, and it is better than Tier 1. But Tier 3 is the one that actually gets used every time, because it accounts for the constraint that matters most and gives you something ready to act on instead of something you have to research further.

Try It Yourself

1. Take the prompt you brought from your own life or work today.
2. Which of the five building blocks does it already have?
3. Rewrite it one tier stronger - even if you can't hit all five yet, add just one you're missing and re-run it.
4. Read the new output. Does it feel more “yours” and less generic?

Garbage in, garbage out isn't a criticism of the AI - it's a reminder that you're always the one steering.

Before Lecture 1.5

Today Is All Hands-On

No new theory today - we're closing out Module 1 by putting everything together. You've learned the mindset shift (Lecture 1.2), the five building blocks of a prompt (Lecture 1.3), and the golden rule of input quality (Lecture 1.4). Today, we turn all three into a repeatable rewrite process you can use on literally any weak prompt, forever.

The Rewrite Process (Use This Every Time)

• Step 1 - Identify the real task. What do you actually need, underneath the vague request?
• Step 2 - Add a Role. Who has the exact expertise to do this well?
• Step 3 - Add Context. What's the real situation the AI needs to know?
• Step 4 - Add Format. How do you want the answer shaped so it's immediately usable?
• Step 5 - Add Constraints. What should it avoid, skip, or limit?

Live Walkthrough 1 - Watch This Get Rebuilt Step by Step

Notice how each step didn't replace the last one - it stacked on top. That's the pattern for the entire rest of this course: every new technique we learn gets added on top of what you already know, not swapped in instead of it.

Live Walkthrough 2 - A Customer Apology Email

Same stacking pattern, a completely different task. Role and Context went in together this time since they were short enough to combine, then Format and Constraints closed it out.

Live Walkthrough 3 - A Product Description

Three different tasks, LinkedIn post, apology email, product description, and the exact same 5-step stacking process built all three. That's not a coincidence, that's the whole point of this lecture.

Your Turn - In-Class Challenge

1. Take the real prompt you've been carrying since Lecture 1.4.
2. Run through all 5 steps out loud or on paper before typing anything.
3. Type the final version into your AI tool.
4. Pair up with one other student (or read it back to yourself) and ask: does this pass the “could this apply to anyone” test from Lecture 1.2?

Module 1 Recap

• AI follows instructions - it doesn't search. Thin input gets thin output.
• Every strong prompt is built from five blocks: Role, Context, Task, Format, Constraints.
• Garbage in, garbage out isn't a limitation of the AI - it's your steering wheel.
• You now have a repeatable 5-step rewrite process you can use on any weak prompt, starting today.

You just finished Module 1. Everything from here is you getting sharper and faster at the exact same five moves.

Before Lecture 2.1

Let's Put a Name on What We Already Did

Back in Lecture 1.2, we typed "best gym marketing ideas" into an AI tool and watched it give us a generic answer. That was your first real example of a technique that actually has a name: zero-shot prompting. Today we're putting a label on it, understanding exactly what it is, and figuring out when it's genuinely the right tool for the job.

The Definition

In prompt engineering, a "shot" means an example. Zero-shot prompting means you're giving the AI a task with zero examples and zero extra context beyond the bare instruction itself. You just ask, and you expect an answer, the same way you'd ask a stranger a quick question on the street.

Example 1: A Poem

Nothing here tells the AI who it's writing for, what mood you want, how long it should be, or what style you like. So it fills every one of those gaps with its own best guess, the most "average" version of an ocean poem it can produce.

Example 2: Interview Advice

Same problem, different task. It doesn't know what job, what industry, what you're nervous about, or what stage of the process you're in, so it gives you the generic advice that applies to literally every interview ever.

Example 3: A Bakery Logo

That's not a flaw in the AI. That's just what happens when you don't specify anything, you get the statistical middle of the road, whether you're asking for a poem, advice, or a design concept.

Zero-Shot Isn't Bad. It's a Starting Point.

I want to be really clear about something: zero-shot prompting is not a mistake. Every single person in this room started here, and you'll still use it every day for simple, low-stakes tasks. The problem only shows up when you use zero-shot on a task that actually needed more direction, and then you're disappointed with an answer that was never going to be specific in the first place.

Try It Yourself

1. Write 3 zero-shot prompts right now, no role, no context, no format, just the bare task. Pick 3 different types of tasks (writing, explaining, brainstorming).
2. Run all 3. Read each output and ask: "could this apply to anyone, anywhere, with zero changes?"
3. Notice which ones felt "good enough" as zero-shot and which ones clearly needed more.

Before Lecture 2.2

Every Technique Has a Ceiling

Zero-shot prompting works great for some tasks and falls apart on others. Today we're learning to tell the difference before you waste time on a prompt that was never going to work.

Where It Holds Up Just Fine

All three are fine as zero-shot. Each has one correct answer, no personal voice involved, no format decisions to make, and no audience to think about. Zero-shot is perfect for facts, quick lookups, and simple, single-answer questions.

Where It Breaks Down

All three of these have no single correct answer, only better and worse fits. Each needs a voice, a specific situation, and a sense of what "good" looks like. Zero-shot can't supply any of that, so it invents a generic, safe, forgettable version instead.

The Complexity Test

Before you write a prompt, ask yourself one question: does this task have one right answer, or does it depend on a personal voice, a specific audience, or a particular format? If it's the first, zero-shot is fine. If it's the second, you need at least a role and some context, which we're covering starting next module.

Try It Yourself

1. Sort these into "zero-shot is fine" and "zero-shot will break": converting 10 miles to kilometers, writing a wedding toast, explaining what a noun is, writing a cover letter, listing synonyms for "happy," drafting a customer apology email.
2. For each one you marked "will break," say out loud what's missing: a voice? A specific audience? A format?

Before Lecture 2.3

When Zero-Shot Is Actually Your Best Friend

We spent last lecture talking about where zero-shot breaks down. Today we're flipping it around: there's a whole category of everyday tasks where zero-shot isn't just "acceptable," it's genuinely the smartest choice, because writing a longer, more detailed prompt would just waste your time for no extra benefit.

Four Everyday Zero-Shot Wins

None of these need a role, a brand voice, or a specific format. They have low stakes, and "good enough" genuinely is good enough. Adding a role like "you are a senior etymologist with 20 years of experience" to a simple definition request would just be wasted typing.

Speed Over Precision

The skill here isn't writing fancier prompts, it's recognizing which tasks don't need one. Save your five-building-block, fully-engineered prompts for tasks where precision actually matters: your brand, your career, a client deliverable. For the small stuff, zero-shot keeps you moving fast.

Try It Yourself

1. List 5 tasks from your own week that are genuinely low-stakes and quick: conversions, definitions, quick rewrites, simple lists.
2. Run all 5 as plain zero-shot prompts, no role, no context.
3. Confirm: did any of them actually need more detail, or were they genuinely fine as-is?

Before Lecture 2.4

Training Your Eye to Spot "Generic"

You already know the test: could this exact output work for a totally different person or business with zero changes? Today we're not just learning the test, we're practicing it until it becomes automatic, across three completely different types of content.

Round 1: Email

Round 2: Social Media Caption

Round 3: Business Advice

Notice the pattern across all three: the generic version could be copy-pasted onto any business's website tomorrow. The specific version only makes sense for this exact business, this exact product, this exact problem. That specificity is what a real role, real context, and real constraints buy you, which is exactly where we're headed starting in Module 3.

Try It Yourself

1. Pull up the "just okay" output you brought from last lecture.
2. Run the anywhere test on it out loud.
3. Identify exactly what's missing: a name, a number, a specific detail, a real situation?
4. Rewrite the original prompt to include that missing piece, and compare the two outputs.

Before Lecture 2.5

Module 2 Speed Round

No new theory today. This lecture is entirely hands-on, four rapid-fire rounds, back to back, to lock in everything from this module before we move on to Role Prompting next.

Round 1: Write It Zero-Shot

Take the task you brought from last lecture. Write it as a bare zero-shot prompt, no role, no context, nothing extra. Run it. Read the output out loud to yourself.

Round 2: Run the Anywhere Test

Ask the question: could this output work for a totally different person with zero changes? Answer honestly, yes or no.

Round 3: Decide, Fine or Broken

Using the complexity test from Lecture 2.2: does this task have one correct answer, or does it depend on a voice, an audience, or a format? Decide which bucket your task falls into.

Round 4: Fix It, If It Needs Fixing

If your task passed as genuinely fine for zero-shot, great, you just saved yourself time. If it needs more, add one specific, real detail (a name, a number, a real situation) and rerun it. Compare the before and after.

Module 2 Recap

• Zero-shot prompting means zero examples, zero extra context, just the bare task.
• It's not wrong, it's a starting point that works great for facts, conversions, and low-stakes quick tasks.
• It breaks down on anything needing a voice, a specific audience, or a format, because it has to guess, and it guesses generic.
• The anywhere test ("could this apply to anyone?") is your fastest way to catch a weak zero-shot output.

You now know exactly when to reach for zero-shot and when to reach for something stronger. That's the whole skill.

Before Lecture 3.1

The First Real Fix for "Generic"

In Module 2, we learned to spot generic output and we learned exactly why it happens: zero-shot prompts leave gaps, and the AI fills those gaps with the most average answer possible. Today we're learning the first technique that actually closes that gap. It's called role prompting, and it might be the single highest-impact thing you learn in this entire course.

The Definition

Role prompting means assigning the AI an identity, a specific expertise, and a point of view before you give it the task. Instead of asking a nameless, faceless assistant for help, you're asking a specific kind of expert, and that expert brings their own vocabulary, priorities, and instincts to the answer.

Watch This Difference

Three tasks, same pattern every time.

Example 1: Business Plan Feedback

The first version gets you polite, generic feedback, the kind you could get from anyone. The second version gets you feedback filtered through a very specific lens: someone whose entire job is deciding who gets money and who doesn't.

Example 2: Resume Review

Example 3: Restaurant Menu Feedback

Notice the pattern across all three: without a role, the AI gives balanced, generalized feedback anyone could offer. With a role, the feedback gets filtered through a specific lens, what an underwriter would flag, what a recruiter would skip past, what a menu consultant would price differently. That lens changes what gets noticed, what gets flagged, and how directly it gets said.

Why This Works

Without a role, the AI has no perspective to draw from, so it gives you a balanced, safe, generalized answer. The moment you assign a role, you activate a specific set of patterns: the vocabulary that role would use, the priorities that role would care about, and the things that role would notice that a generalist would miss entirely.

Try It Yourself

1. Take a task you'd normally ask for help with (feedback, advice, a review, a plan).
2. Run it once with no role at all.
3. Run it again, this time starting with "You are a [specific expert] with [specific experience]."
4. Compare: which one actually sounds like it came from someone who does this for a living?

Before Lecture 3.2

A Title Is Not a Lens

Here's a mistake almost everyone makes when they first learn role prompting: they think just naming a job title is enough. "You are a doctor." "You are a teacher." "You are a lawyer." These are titles, and titles alone barely move the needle. What actually works is a lens, a specific angle, specific experience, and a specific thing that role cares about.

Weak Role vs. Strong Role, Three Jobs

Job 1: Doctor

"You are a doctor" is weak because there are a thousand kinds of doctors with a thousand different instincts. "An ER doctor who decides fast under pressure and talks to scared patients" is a lens.

Job 2: Teacher

Job 3: Lawyer

Same pattern, three completely different professions. Each weak role is just a title, and each strong role names a specific angle, specific experience, and a specific thing that person always notices. It tells the AI exactly what to prioritize, not just what field to sound like it's from.

Three Ingredients of a Strong Role

• Specific expertise, not just a title ("senior financial analyst who audits startup budgets," not just "analyst")
• Real experience, years or a specific kind of work ("15 years underwriting small business loans")
• A known priority or pet peeve, what that person always notices or always cares about ("catches hidden burn-rate risk that founders miss")

Try It Yourself

1. Take the role you brought from last lecture.
2. Check it against the three ingredients above. Which ones are missing?
3. Rewrite it to include all three, then rerun your prompt and compare.

Before Lecture 3.3

Role Prompting Is a Tone Dial

One of the most practical uses of role prompting has nothing to do with expertise at all, it's about tone. The exact same task can come out warm, blunt, playful, formal, or reassuring, purely based on the role you assign. Today we're using role prompting as a tone control knob.

Same Task, Four Roles, Four Tones

Same rejection, same news, three completely different feelings behind it. None of these is "wrong," they're just built for different brands and different relationships. This is why brand voice matters so much, and role prompting is the fastest way to control it.

Writing a Tone-Focused Role

When your goal is tone, describe the role using how they talk, not just what they know. "A friendly startup founder who talks like a real person, not a corporation" tells the AI more about tone than "an HR professional" ever could.

Try It Yourself

1. Take the writing task you brought today.
2. Write it with a role focused purely on tone: warm, blunt, playful, formal, whatever fits your situation.
3. Run it, then try one more tone and compare both against what you actually need for this message.

Before Lecture 3.4

Raising the Bar for Technical Work

Tone is one use of role prompting. Today we're looking at a completely different use: getting genuinely sharper, more rigorous answers on technical or expert-level tasks, the kind where a generic answer isn't just boring, it's actually unhelpful or even risky.

Generalist vs. Specialist, Three Domains

Domain 1: Code Review

The first prompt might get you a shrug and "looks fine to me." The second one activates a specific kind of scrutiny, the exact scrutiny a real strict reviewer would apply.

Domain 2: Financial Analysis

Domain 3: Legal Contract Review

Across all three domains, the generic version gets you a shrug. The specialist role activates real scrutiny, because you told the AI exactly what that kind of expert watches for.

Matching the Role to the Real Stakes

For technical and expert tasks, your role should describe the standard you want applied, not just a job title. "A security-focused engineer who assumes every input is malicious until proven otherwise" gives you a very different code review than "a helpful coding assistant." Pick the standard that matches what's actually at stake in your task.

Try It Yourself

1. Take the technical or expert task you brought today.
2. Write a specialist role that names the exact standard or scrutiny you want applied (strict, thorough, security-focused, skeptical, whatever fits).
3. Run it and check: did it catch something a generic "helpful assistant" answer would have missed?

Before Lecture 3.5

Build Your Personal Role Library

This is the lecture where Module 3 turns into something you'll actually use for the rest of your life: a small, reusable set of roles built specifically for the tasks you do regularly. Once you write these, you'll never start from a blank page again.

Three Finished Examples to Work From

Notice all three follow the same three-ingredient shape from Lecture 3.2, just aimed at completely different tasks. That's the whole point, one template, endless uses.

Your Turn: Build 5 Roles

1. Pick 5 tasks you actually do regularly (from work, school, or your personal life).
2. For each one, write a role using the three-ingredient template: expertise, experience, priority or pet peeve.
3. Save all 5 in your notes doc. Label each one clearly with the task it's for, exactly like we labeled few-shot examples back in the roadmap preview.
4. Test at least 2 of them right now with a real task.

Module 3 Recap

• Role prompting means assigning the AI a specific identity and point of view before the task.
• A title alone is weak. A lens (expertise, experience, and a known priority) is strong.
• Roles can control tone (how something sounds) or rigor (how carefully something gets checked).
• A small library of 5 to 10 saved roles will cover almost everything you do on a regular basis.

You just built your first real, reusable prompting asset. Keep this file open, you'll be adding to it for the rest of the course.

Before Lecture 4.1

When Describing Isn't Enough

Role prompting solved a lot of our generic-output problem, but there's a specific wall you'll hit even with a great role: describing a voice in words is genuinely hard. Try describing your own brand's tone in a sentence, or the exact vibe of your favorite writer. It's slippery. Today's technique solves that by skipping the description entirely.

The Definition

Few-shot prompting means giving the AI a small number of labeled examples of exactly the kind of output you want, before asking for a new one. Instead of telling the AI what your voice sounds like, you show it, and it learns the pattern directly from the examples.

Example 1: Coffee Brand Email

The first prompt gives the AI adjectives to interpret: witty, direct, irreverent. It has to guess what that actually sounds like in sentences. The second prompt gives it real sentences to learn the rhythm from.

Example 2: Fitness App Push Notification

Example 3: Law Firm Client Update

Same pattern, three completely different brands. Each role-only version leaves the AI guessing what "witty" or "reassuring" actually sounds like in real sentences. Each few-shot version hands it the actual rhythm to match.

Why This Works

Think about how hard it is to describe a music genre in words versus just playing someone two or three songs. They'll get it instantly. Few-shot prompting works the same way: showing examples encodes a pattern far more precisely than any adjective-filled description ever could.

Try It Yourself

1. Pick a piece of writing you'll need this week (an email, a post, a message).
2. Write it once using only a role description (adjectives, no examples).
3. Find or write 2 real examples of writing in the voice you actually want, and run the same task again with those examples included.
4. Compare: which one actually sounds right?

Before Lecture 4.2

Show, Don't Tell (For AI, Too)

Writers hear "show, don't tell" constantly. Turns out it's just as true for prompting an AI as it is for writing a novel. Today we're looking at why examples carry more information than descriptions, and where this matters most.

The Music Genre Problem

If I asked you to describe the genre "lo-fi hip hop" in one sentence, you'd struggle, and whatever you said would miss something. But if I played you three songs, you'd understand it immediately, and you could probably pick a fourth song from the same genre out of a lineup. That's the entire idea behind few-shot prompting: some things are easier to recognize than to define.

Where This Matters Most

• Brand voice, tone is nearly impossible to fully describe in adjectives
• Code style, showing 2 functions you've written teaches formatting and naming conventions instantly
• Structured formats, showing one filled-out example row teaches a table format faster than describing every column
• Editing style, showing a before-and-after edit teaches your preferences better than a list of rules

Example 1: Coding Style

The description tells the AI what to aim for. The example shows it the actual target: naming pattern, docstring style, type hints, even spacing.

Example 2: Table Format

Example 3: Editing Style

One good example often beats a whole paragraph of rules, whether you're teaching a coding style, a table format, or an editing preference.

Try It Yourself

1. Pick something you personally have a style for: how you write texts, how you format a to-do list, how you title your documents.
2. Try describing that style in a sentence.
3. Then just paste a real example of it and ask the AI to match it.
4. Notice which one actually captured your style better.

Before Lecture 4.3

Not Just Any Examples Will Do

Few-shot prompting only works if your examples are actually good examples. Today we're covering two rules that separate few-shot prompts that work from ones that quietly confuse the AI: labeling and variety.

Rule 1: Label Every Example

Don't just paste examples in a pile. Name each one so the AI understands what kind of example it's looking at. This also helps you organize your own thinking about what range of outputs you're trying to teach.

Rule 2: Show Variety, Not Repetition

If all your examples are the same length, cover the same topic, and follow the exact same structure, the AI will assume that structure is a rule to copy exactly, not just a voice to learn. Show a short one and a long one. Show different topics. The voice should stay consistent even as the content changes.

How Many Examples Do You Need?

Two to five is the sweet spot for almost everything. One example isn't enough to show a pattern versus a coincidence. More than five rarely adds much, and it makes your prompt harder to maintain.

Real World Examples Beat Made-Up Ones

Whenever possible, use examples you've actually written or that genuinely represent your brand, not examples you invent on the spot to sound good. Real examples carry real nuance that made-up ones miss.

Try It Yourself

1. Take the 2 or 3 real examples you brought today.
2. Label each one clearly by type or purpose.
3. Check: are they different enough in length or topic, or do they all look the same? If they're too similar, that's fine for now, just note it.
4. Use them as few-shot examples for a new piece of writing and check the result.

Before Lecture 4.4

When Few-Shot Goes Wrong

Few-shot prompting fails in a very specific, very common way: overfitting. That means the AI copies surface details from your examples, like length, structure, or specific phrases, instead of learning the underlying voice you actually wanted. Today we're learning to spot it and fix it.

What Overfitting Looks Like, Three Cases

Every set here has the exact same opening, the exact same closing, and the exact same structure, just one word swapped. Feed any of these to the AI for a genuinely new task, and it may force in that same rigid shape because that's the only pattern it saw, not because it makes sense for the new task.

The Fix: Deliberate Variety

Go back to Lecture 4.3's rule: vary the length, the topic, and the structure across your examples, while keeping the actual voice consistent. This teaches the AI what stays the same (the personality) and what should change (the specific structure) based on the task.

A Second Common Mistake: Outdated Examples

If your examples reflect an old brand voice, an old job, or a style you've since moved past, the AI will faithfully reproduce that old version. Always double check that your examples still represent where you or your brand actually are right now.

Try It Yourself

1. Look back at the examples you labeled last lecture.
2. Check them for overfitting: do they all open the same way, close the same way, or run the same length?
3. If yes, rewrite or replace one example so it's meaningfully different in structure while keeping the same voice.
4. Rerun your few-shot prompt with the updated set and compare.

Before Lecture 4.5

Train the AI on Your Own Voice

This is the lecture where Module 4 becomes something you'll use constantly. We're taking everything from this module, definitions, why showing beats telling, choosing good examples, avoiding overfitting, and building one complete, reusable voice-matching prompt for yourself.

The Full Recipe

• Step 1, gather 2 to 5 real examples of your own writing (from Lecture 4.2 and 18).
• Step 2, label each one clearly by type or purpose.
• Step 3, check for variety, different lengths and topics, same underlying voice (Lecture 4.4).
• Step 4, add a short role description on top for extra clarity (from Module 3).
• Step 5, ask for a brand new piece of writing in that voice.

Three Finished Examples

Notice all three follow the identical shape, a short voice description, two labeled real examples, then the new task. Only the voice and the task change, the structure of the prompt itself stays the same every time.

Your Turn

1. Build your own version of the prompt above, using your real examples from today.
2. Run it on a brand new task you haven't tried yet.
3. Read the result out loud. Does it sound like you, or like a generic assistant wearing your name?
4. Save this finished prompt template in your notes doc, this is now a permanent tool in your library.

Module 4 Recap

• Few-shot prompting means showing labeled examples instead of describing a style in words.
• Showing beats telling because some things are easier to recognize than to define.
• Good examples are labeled and varied, not repetitive or outdated.
• Overfitting happens when examples are too similar, the AI copies structure instead of learning voice.

You can now make an AI sound like you, your brand, or anyone else whose voice you can show it. That's a genuinely rare skill.

Before Lecture 5.1

The Wall of Text Problem

Raise your hand if this has happened to you: you ask an AI for something, and you get back one giant, unbroken paragraph, or a huge dump of information with no organization at all, and you're left manually pulling it apart to find what you need. That's what today fixes. Structured prompting means you tell the AI exactly how the answer should be shaped, not just what to say.

The Definition

Structured prompting means explicitly defining the format of the response: sections, word or character limits, tables, labeled parts. You're not just asking a question, you're handing over a shape for the answer to fill.

Example 1: Multi-Platform Marketing

The unstructured version gives you a wall of mixed content you have to manually separate before you can use any of it. The structured version hands you three ready-to-use pieces, already labeled.

Example 2: Code Review Request

Example 3: Job Description Request

Same lesson, three different tasks. In every case, the structured version hands you something you could use immediately, and the unstructured version hands you something you'd have to reformat by hand before you could use it at all.

Why This Matters

Structure isn't about limiting the AI's intelligence. It's about not making your future self do the formatting work by hand. A great answer that's shaped wrong still costs you time. A structured answer is usable the moment it lands.

Try It Yourself

1. Think of a task where you got a big blob of text back recently.
2. Rerun that exact task, but this time tell the AI the structure you want: sections, a list, a specific order.
3. Compare how much editing each version needs before you could actually use it.

Before Lecture 5.2

Designing the Shape You Actually Need

Now that you know structured prompting exists, let's learn how to actually design a format. There are three tools you'll use constantly: sections, limits, and tables.

Tool 1: Sections

Break the answer into clearly labeled parts, so each piece is easy to find and use on its own.

Tool 2: Limits

Give a real number, words, characters, or items, not a vague instruction like "keep it short." Vague words get vague results. Numbers get exact results.

Tool 3: Tables

When you're comparing multiple things across the same categories, ask for a table with named columns. This is one of the clearest, fastest wins in all of structured prompting.

Across all three tools, the pattern is the same: a real structure, named parts, real numbers, gives you an answer you can use immediately, not one you have to reshape by hand first.

Try It Yourself

1. Take the task you brought from last lecture.
2. Decide which tool fits best: sections, limits, or a table (you can use more than one at once).
3. Write the full structured prompt and run it.
4. Check: is the output usable immediately, with zero manual reformatting?

Before Lecture 5.3

One Prompt, Many Platforms

Here's where structured prompting really pays off: instead of running three or four separate prompts for email, Instagram, Twitter, and your website, you can generate all of them in one shot, each one correctly sized and shaped for its platform.

Example 1: A Bakery Launch

Notice how each platform gets its own rules. Email gets a subject line limit, Instagram gets hashtag guidance, Twitter gets character limits per post. One prompt, four ready-to-use deliverables.

Example 2: A Fitness App Feature Launch

Example 3: A Nonprofit Fundraiser

Same structuring principle, three completely different purposes. Once you know the platform rules, swapping the topic and tone takes seconds.

Why Platform-Specific Rules Matter

A caption that works on Instagram would look strange as a tweet. An email subject line rule doesn't apply to a hashtag. If you don't specify platform-specific limits, the AI will average everything into one generic length that fits none of them particularly well.

Try It Yourself

1. Take the content idea you brought today.
2. Build a multi-platform structured prompt covering at least 2 platforms (email, Instagram, Twitter, LinkedIn, whatever fits your situation).
3. Give each platform its own specific limits.
4. Run it once and check: could you publish each piece immediately, with no edits?

Before Lecture 5.4

Structure for Serious Documents

Multi-platform content is one use of structure. Today's use is more formal: reports and comparisons, the kind of document someone actually has to read, act on, or forward to a boss or a client. These need a different kind of structure, one built for clarity and scannability.

Three Report Structures

Three Comparison Structures

All six of these give the reader a predictable shape to scan. A busy executive can jump straight to "Next Steps" or straight to "Recommendation" without reading a single unnecessary sentence.

The Golden Rule for Report Structure

Design your structure around what your reader needs to find fastest, not around what feels most complete to write. A report that's thorough but disorganized helps no one. A report that's short but perfectly organized gets read and used.

Try It Yourself

1. Take the comparison task you brought today.
2. Design a structure with named sections or table columns, matching what your reader actually needs to see first.
3. Run it and check: could someone skim this in 15 seconds and know exactly what to do next?

Before Lecture 5.5

Build a Template You'll Use Forever

This is the lecture where Module 5 turns permanent. If you create the same type of content on a recurring basis, a weekly report, a regular content batch, a comparison you run every quarter, you should never have to redesign the structure from scratch again. Today we're building that template once, properly, so you can just fill in new details every time.

The Template-Building Process

• Step 1, identify what you create repeatedly.
• Step 2, list exactly what pieces of information always need to appear.
• Step 3, decide the shape for each piece: a section, a limit, or a table.
• Step 4, write the structure once as a reusable block.
• Step 5, save it, and just swap in new details each time you use it.

Three Finished, Reusable Templates

Notice the only thing that changes in each template is the bracketed placeholder. Everything else, the structure, the limits, the sections, stays exactly the same. That consistency is what makes a template a template.

Your Turn

1. Take the recurring content type you brought today.
2. Build your own reusable structured template using the 5-step process above.
3. Test it twice with two different topics or details, keeping the structure identical both times.
4. Save the finished template in your notes doc, clearly labeled.

Module 5 Recap

• Structured prompting means defining the shape of the answer, not just the content.
• Sections, limits, and tables are your three main structuring tools.
• Multi-platform prompts can generate several ready-to-use pieces in a single request.
• Recurring tasks deserve a permanent, reusable template, not a rebuilt prompt every time.

You just eliminated an entire category of manual formatting work from your life, for good.

Before Lecture 6.1

The Most Powerful Technique in This Course

Everything we've covered so far, roles, examples, structure, has controlled how an answer sounds and how it's shaped. Today's technique is different: it controls how the AI actually thinks through a problem before answering. If I had to pick one technique to teach twice, it would be this one.

The Definition

Chain-of-thought prompting means instructing the AI to reason through a problem step by step, out loud, instead of jumping straight to a conclusion. You're not just asking for an answer, you're asking to see the thinking that gets there.

Practice Document for This Module

Example 1: The AuraTech Buds Pro Campaign

Run both of these against the real report. The first prompt will get you an answer, but it'll almost always be the single most obvious guess. The second forces the AI to actually examine the evidence in order before committing to a conclusion.

Example 2: An Online Store's Sales Drop

Example 3: A Spike in Employee Turnover

Three completely different situations, a marketing campaign, a retail store, an HR problem, and the exact same fix works every time. That's the pattern to notice.

Would You Trust a One-Second Diagnosis?

Imagine a doctor who gives you a diagnosis in one second without asking a single question. You wouldn't trust it, no matter how confident they sounded. Without chain-of-thought, that's exactly what you're getting from an AI on any hard question: instant pattern matching, dressed up as an answer.

Try It Yourself

1. Upload the AuraTech Buds Pro report and run Example 1's straight question, then the chain-of-thought version.
2. Pick a second real question from your own life or work.
3. Ask it once with no reasoning steps, then again with "think through this step by step" and 3 or 4 numbered steps.
4. Compare: which answer would you actually trust more?

Before Lecture 6.2

Four Words That Change Everything

Today we're focusing on the simplest possible version of chain-of-thought: just adding the phrase "think step by step" to a prompt, with nothing else added. It sounds too simple to matter. It matters more than almost anything else in this course.

Example 1: The AuraTech Buds Pro Campaign

That's the entire difference in wording, four words. But it changes the AI's whole approach on a real report: instead of pattern-matching to the most common explanation, it slows down and works through the evidence piece by piece.

Example 2: Quiz Score Results

Example 3: Website Bounce Rate

It Doesn't Need to Be Fancy

You don't need special formatting, capital letters, or exact wording. "Walk me through your reasoning," "reason through this before answering," and "think step by step" all trigger the same effect. Pick whichever phrase feels natural to you and use it consistently.

When to Reach for It

• Anything analytical, diagnosing a problem, interpreting data, figuring out a root cause
• Anything with multiple plausible explanations, where jumping to the first guess would be a mistake
• Anything you'd want a real expert to actually think through, not just answer instantly

You don't need chain-of-thought for simple factual questions, that's still a job for zero-shot from Module 2. This technique earns its keep specifically on harder, more ambiguous problems.

Try It Yourself

1. Run Example 1 on the actual AuraTech report, both with and without the phrase.
2. Pick one simple fact, one analytical question, and one genuinely hard problem of your own.
3. Add "think step by step" to all 3 and run them.
4. Notice: did it help all three equally, or mostly the harder ones?

Before Lecture 6.3

From Magic Phrase to Real Diagnostic Steps

"Think step by step" alone is powerful, but for genuinely complex problems, you can go further: write the actual steps you want the AI to follow, in order. This is like being the one who designs the checklist, instead of just asking someone to "be thorough."

Example 1: The AuraTech Buds Pro Campaign, Full Diagnostic Chain

Run this directly against the uploaded report. Notice each step points at a real section of the document, weekly trends, competitor activity, channel breakdown, audience segments, so the AI has to genuinely use the report, not just summarize it.

Example 2: A Customer Churn Diagnosis

Example 3: An App Store Rating Decline

How to Write Good Steps

• Move from broad to specific, overall trend first, then narrow in
• Each step should answer one clear question, not several at once
• The last step should always be a synthesis step, connect everything found in the earlier steps
• Ask it to show its reasoning at each step, not just the final answer

Try It Yourself

1. Run Example 1 against the actual AuraTech report and read every step of the response.
2. Take a messy problem of your own and write 4 or 5 diagnostic steps, broad to specific, ending with a synthesis step.
3. Check: did the final conclusion actually follow logically from the earlier steps?

Before Lecture 6.4

Where This Technique Pays Off Most

Chain-of-thought prompting earns its keep the most on real data and real business problems, the kind with numbers, trends, and multiple possible explanations. Today we're going one level deeper into the same AuraTech report, proving one document can support several different chain-of-thought investigations, not just one.

Example 1: Why Did Instagram Underperform?

Notice this asks a completely different question than Lecture 6.3's example, but pulls from the exact same uploaded report. That's the real skill: one good source document can answer many different questions if you ask chain-of-thought questions on purpose.

Example 2: A Drop in Email Open Rate

Example 3: A Spike in Support Ticket Volume

All three examples follow the identical broad-to-specific, synthesis-at-the-end shape, whether you're working from a real campaign report or a support ticket count.

A Word of Caution

Chain-of-thought makes the AI reason more, but reasoning confidently isn't the same as reasoning correctly. On real business data, always sanity-check the final conclusion against the actual numbers yourself, especially before making a real decision based on it. We'll cover how to prevent fabricated numbers specifically in Module 7.

Try It Yourself

1. Run Example 1 against the real AuraTech report and check the reasoning against the actual channel and creative data yourself.
2. Pick one more real dataset or report of your own, or keep using AuraTech, and write a full diagnostic chain-of-thought prompt asking a question we haven't asked yet.

Before Lecture 6.5

Full Diagnosis, Start to Finish

This is the hands-on close to Module 6. No new theory today, we're taking a real problem all the way from a vague question to a reasoned, step-by-step diagnosis, using everything from this module together.

The Full Process

• Step 1, state the problem clearly, including any specific numbers you already know (from Lecture 6.1).
• Step 2, add a role that matches the kind of expert who'd normally diagnose this (from Module 3).
• Step 3, write 4 to 5 diagnostic steps, broad to specific (from Lecture 6.3).
• Step 4, end with a synthesis step asking for the most likely cause and the reasoning behind it.
• Step 5, sanity-check the final answer against what you actually know to be true.

Walkthrough 1: The Full AuraTech Diagnosis

Run this one against the real report right now. This pulls together nearly every section of the document, weekly trends, competitor activity, channels, creatives, and the checkout funnel, into one complete diagnosis.

Walkthrough 2: A Delivery Delay Diagnosis

Walkthrough 3: A Restaurant No-Show Rate Spike

Your Turn

1. Run Walkthrough 1 against the real AuraTech report.
2. Pick one problem from your own life or work and build your own complete chain-of-thought prompt using all 5 steps.
3. Run it, read every step of the reasoning, not just the final answer.
4. Decide: do you trust this conclusion? Why or why not?

Module 6 Recap

• Chain-of-thought prompting means asking the AI to reason through a problem step by step instead of jumping to a conclusion.
• "Think step by step" alone is powerful. Writing your own specific steps is even more powerful.
• Good steps move from broad to specific and end with a synthesis step.
• One real source document, like the AuraTech report, can support many different chain-of-thought investigations, not just one.
• This technique makes reasoning more thorough, not automatically correct, always sanity-check real conclusions.

You now know how to make an AI actually investigate a problem instead of just guessing at it. That's a genuinely rare skill to have.

Before Lecture 7.1

Telling It What Not to Do

Last module ended with a warning: chain-of-thought makes the AI reason more, but reasoning confidently isn't the same as reasoning correctly. Today we're fixing exactly that problem with a technique that's the mirror image of everything we've done so far. Instead of telling the AI what to do, we're telling it what not to do.

The Definition

Constraint-based prompting, also called negative prompting or guardrailing, means explicitly telling the AI what to avoid, what to limit, and what not to include. It's the steering that keeps a strong prompt from wandering off in the wrong direction.

Example 1: The AuraTech Buds Pro Campaign

Try both against the real report. The unconstrained version might casually mention something like "the average fitness wearable CPA is X," a number pulled from nowhere. The constrained version either sticks strictly to the report or honestly says it doesn't have that outside information.

Example 2: Customer Churn

Example 3: Website Traffic Drop

Why This Matters More as Prompts Get Stronger

Here's something worth noticing: the stronger and more open-ended a prompt is, the more room the AI has to wander, add unrequested advice, invent a supporting detail, or run far longer than you needed. Constraints don't fight against your other techniques, they protect the results your other techniques worked to produce.

Try It Yourself

1. Run Example 1 against the real AuraTech report, with and without constraints.
2. Take a chain-of-thought prompt of your own from Module 6.
3. Add one constraint telling it to only use the data you provided, and flag anything else as a gap. Rerun and compare.

Before Lecture 7.2

The Fabrication Problem

AI models can state a made-up fact with exactly the same confident tone as a real one. This is often called hallucination, and it's the single biggest reason you should never fully trust an AI's numbers or citations without checking. Today we're learning the specific constraints that reduce this risk.

The Anti-Fabrication Constraint, Three Versions

Same core constraint, three different documents. It does two things at once every time: it restricts the AI to your actual source material, and it gives it explicit permission to say "I don't have that information" instead of guessing.

The Commit-or-Flag Constraint, Three Versions

This prevents vague hedging that quietly smuggles in unsupported claims, whether you're diagnosing a campaign, summarizing symptoms, or reviewing a candidate.

Try It Yourself

1. Take the output you brought from last lecture.
2. Identify any fact, number, or claim you can't verify against your original source.
3. Rerun the original prompt with both constraints added, the anti-fabrication rule and the commit-or-flag rule.
4. Compare: did it remove the questionable claim, or flag it honestly instead?

Before Lecture 7.3

Three More Things Worth Constraining

Fabrication isn't the only thing constraints fix. Today we're covering three more common problems: answers that run too long, tone that drifts from what you needed, and scope that wanders into territory you never asked about.

Constraining Length, Three Examples

Constraining Tone, Three Examples

Constraining Scope, Three Examples

Without scope constraints, chain-of-thought prompts especially tend to drift: the AI diagnoses the problem, then keeps going and starts recommending unrelated fixes you never asked for.

Try It Yourself

1. Take the too-long or off-topic output you brought today.
2. Identify which constraint it's missing: length, tone, or scope.
3. Add that constraint, with a short reason attached, and rerun the prompt.
4. Compare the before and after.

Before Lecture 7.4

A Mindset Shift About Constraints

Some people hesitate to add constraints because it feels like they're limiting the AI, holding back its full ability. Today's lecture is about correcting that instinct completely, because it's backwards.

The Steering Wheel Metaphor

A steering wheel doesn't make a car weaker or slower. It's the thing that lets the car go anywhere useful at all, instead of driving straight off the road. Constraints work the same way: they don't reduce the AI's intelligence, they aim it. Every constraint you add is one less wrong direction the AI is free to wander into.

Example 1: The AuraTech Report

Example 2: A Code Review Request

Example 3: A Draft Email

Read each guardrailed version again. None of them is smaller or less capable than its unconstrained counterpart, each one is precisely aimed. Every constraint removed one way the answer could have gone wrong.

A Simple Rule of Thumb

Anytime an AI output disappoints you in a specific, repeatable way, too long, too vague, mentions something unverifiable, wanders off topic, don't just accept it as a limitation. Turn that specific disappointment into a constraint, and it usually won't happen again.

Try It Yourself

1. Take the prompt you brought today.
2. List every way the output could realistically go wrong: too long, off-topic, invents facts, wrong tone.
3. Write one constraint for each risk you listed.
4. Rerun the fully guardrailed version and compare it to the original.

Before Lecture 7.5

Guardrail a Prompt That Already Works

This is the hands-on close to Module 7. We're taking a prompt you already like, one that already produces good output, and making it more reliable by adding the right constraints. This is exactly the habit I want you to build for every strong prompt you write from now on.

The Guardrail Checklist

• Fabrication check, could it invent a fact, number, or source? Add the anti-fabrication constraint from Lecture 7.2 if so.
• Length check, could it run longer than useful? Add a limit with a reason.
• Tone check, could it drift into a tone you don't want? Name what to avoid.
• Scope check, could it wander into unrequested advice or topics? Tell it to stay focused.
• Commitment check, could it hedge instead of giving a clear answer? Ask it to commit or flag insufficient evidence.

Example 1: The AuraTech Campaign Diagnosis

Example 2: A Budget Review

Example 3: A Content Calendar Request

Your Turn

1. Take the already-good prompt you brought today.
2. Run it through the full guardrail checklist above.
3. Add every constraint that applies.
4. Run both versions and compare, is the guardrailed one just as good, but more reliable?

Module 7 Recap

• Constraint-based prompting means telling the AI what to avoid, not just what to do.
• The strongest use is preventing fabrication, restrict the AI to your real source, and let it flag gaps instead of guessing.
• Length, tone, and scope are all constrainable, and pairing a limit with a reason makes it more effective.
• Constraints are a steering wheel, not a leash, they aim your strongest prompts instead of weakening them.

From here forward, treat guardrailing as a standard last step on every important prompt you write, not an optional extra.

Before Lecture 8.1

Everything You've Learned, One Name

Way back in Lecture 1.3, we broke a prompt down into five building blocks: role, context, task, format, and constraints. Since then, you've learned an entire technique for each one. Today we're bundling all five into a single, reusable framework with a name you'll never forget: CRAFT.

The Definition

• C, Context, the real background or situation
• R, Role, who the AI should act as (Module 3)
• A, Action, exactly what you want it to do
• F, Format, how the output should be shaped (Module 5)
• T, plus Constraints, what it should avoid (Module 7)

CRAFT isn't a new technique you're learning from scratch. It's a checklist that makes sure you never forget one of the pieces you already know how to write. Think of it as the recipe card that holds everything else you've learned. Over this module, we're going to build up 3 running examples together, a candle business loan pitch, a fitness studio's marketing plan, and a freelancer's contract review, watching all three go from a bare idea to a complete CRAFT prompt.

Example 1: A Loan Pitch

Example 2: A Fitness Studio Marketing Plan

Example 3: A Freelance Contract Review

Notice all three are just five labeled sections. Nothing here is new information, it's Lecture 3.2's role, Lecture 5.1's format thinking, and Lecture 7.1's constraints, all organized under one clear structure, no matter the domain.

Why Bother Naming It?

Once something has a name, it's easy to remember and easy to reach for under pressure. When you're staring at a blank prompt box, "CRAFT" is a five-second mental checklist: did I cover context, role, action, format, and constraints? If yes, you're almost certainly writing a strong prompt.

Try It Yourself

1. Pick any task you need help with today.
2. Write it out labeled exactly like the examples above: CONTEXT, ROLE, ACTION, FORMAT, CONSTRAINTS.
3. Run it, then read back through and confirm you actually filled in all five.

Before Lecture 8.2

Setting the Scene Properly

Context and Role are the two blocks that come first in CRAFT for a reason: they set the scene before any task even gets mentioned. Today we're focused entirely on writing these two well, since a weak Context or Role quietly weakens everything that follows. We'll rebuild all 3 running examples from Lecture 8.1, starting with just these two blocks.

Writing a Strong Context

Context answers one question: what's the real situation? Not the task itself, the background that explains why the task matters and what's actually at stake.

Across all three, the weak version tells the AI almost nothing about stakes or purpose. The strong version explains exactly why the task matters and what "good" looks like.

Writing a Strong Role

This is everything from Module 3: specific expertise, real experience, and a known priority, not just a job title.

Try It Yourself

1. Take the task you brought today.
2. Write a strong Context: the real situation and what's at stake.
3. Write a strong Role: specific expertise, experience, and a known priority.
4. Notice how much clearer the task feels already, even before adding Action, Format, or Constraints.

Before Lecture 8.3

Now Tell It What to Actually Do

Context and Role set the scene. Action and Format tell the AI exactly what to produce and exactly how it should look. Today we're finishing the picture from last lecture, for all 3 running examples.

Writing a Strong Action

Action should be specific and outcome-focused, not vague. Instead of "help me with this," name the exact deliverable you want.

Writing a Strong Format

This is Module 5 in miniature: sections, limits, or tables, whatever shape makes the answer immediately usable.

Putting Four Blocks Together

Four of five blocks down, across all three examples. Notice how specific and usable each one already feels, even before we add the final piece, Constraints, next lecture.

Try It Yourself

1. Take your Context and Role from last lecture.
2. Add a strong, specific Action naming the exact deliverable.
3. Add a strong Format defining the shape of the answer.
4. Run all four blocks together and check the result.

Before Lecture 8.4

The Final Piece, and Making It Reusable

Today we finish CRAFT with Constraints, the guardrails from Module 7, and then take the completed framework one step further: turning it into a template you can reuse again and again, not a one-time prompt you write from scratch every time.

Three Complete Five-Block CRAFT Prompts

All three are now complete CRAFT prompts, all five blocks, nothing missing. This is the shape every serious prompt in your library should aim for from now on.

Turning Each One Into a Reusable Template

The trick to reusability is separating what stays the same from what changes each time. Context and Action often need small edits per use. Role, Format, and Constraints frequently stay identical across many uses of the same template.

Try It Yourself

1. Take your four-block prompt from last lecture.
2. Add Constraints: what should it avoid, and what's the length limit?
3. Mark which blocks would stay the same next time you use this prompt, and which would need to change.
4. Save it as a labeled template in your notes doc.

Before Lecture 8.5

Build Your Personal CRAFT Library

This is the lecture where Module 8 becomes something you'll use for the rest of your life. The best prompt engineers aren't the ones who write brilliant prompts from scratch every time, they're the ones with a small folder of CRAFT templates they just tweak in 30 seconds. Today, you're building that folder, and we're branching into 3 brand new domains to prove this works for absolutely anything.

The Build Process

• Step 1, list 3 to 5 tasks you do on a recurring basis.
• Step 2, for each one, write all five CRAFT blocks in full.
• Step 3, mark which blocks change each time (usually Context and Action) and which stay fixed (usually Role, Format, and Constraints).
• Step 4, save each template clearly labeled by task.
• Step 5, test each one on a real task right now.

Three More Full Examples, Brand New Domains

Your Turn

1. Build 3 to 5 complete CRAFT templates for your own recurring tasks.
2. Clearly label each one by what it's for.
3. Test at least 2 of them on real tasks today.
4. Save all of them together in one place in your notes doc, this is officially your permanent prompt library.

Module 8 Recap

• CRAFT bundles everything you've learned so far into one checklist: Context, Role, Action, Format, Constraints.
• Context and Role set the scene. Action and Format define the deliverable. Constraints protect the result.
• A CRAFT prompt becomes a reusable template once you know which blocks change each time and which stay fixed.
• A small library of 3 to 5 templates covers most of what you'll ever need to prompt for.

You now have a repeatable system, not just a set of tricks. That's the real turning point in this course.

Before Lecture 9.1

When One Prompt Isn't Enough

You now have CRAFT, a complete framework for writing a strong single prompt. But some jobs are simply too big for one prompt, no matter how well you write it. A long report, a full meeting transcript, a messy research question. Today we're learning what to do when the job itself needs to be broken apart, not just the prompt.

Practice Documents for This Module

The Definition

Task decomposition means breaking a big job into a sequence of smaller prompts, where each response builds on the last, instead of writing one giant prompt and hoping for the best. Each prompt hands context forward to the next one.

Example 1: The AuraTech Campaign Report

Example 2: The SkillBridge Meeting Transcript

Example 3: Customer Survey Results

In every case, the mega-prompt asks for everything at once, and the AI has to guess how deep to go on each part, often skimming all of them. The chain gives each step room to be genuinely thorough, and each step gets smarter because it's building on real analysis, not starting cold.

The House-Building Analogy

This is the difference between saying "build me a house" all at once versus saying "first the foundation, then the frame, then the walls, then the roof." Same house, much better process, and far less chance something gets skipped.

Try It Yourself

1. Run Example 1's mega-prompt against the real AuraTech report, then imagine running the 3-prompt chain instead.
2. Think of a big task of your own you've been avoiding because it feels overwhelming to prompt all at once.
3. Write down 2 or 3 smaller stages that job could be broken into.

Before Lecture 9.2

Building the Chain, Link by Link

Today we're taking the idea from last lecture and turning it into real, working sequences of prompts. The key skill here isn't writing any single prompt, it's deciding what belongs in which link of the chain.

Chain 1: The AuraTech Campaign Report

Chain 2: The SkillBridge Meeting Transcript

Notice that Prompt 2 never repeats the raw data in either chain, it just says "the 3 buckets you just created" or "the action items you just extracted," trusting the conversation history to carry that context forward. Each prompt is short precisely because it isn't starting from zero.

Why This Also Makes You Sharper

There's a side benefit here that's easy to miss: reading each response before writing the next prompt makes you ask better questions. By the time you write Prompt 3, you actually understand the situation well enough to ask something genuinely useful, not just generic.

Try It Yourself

1. Run Chain 1 against the real AuraTech report, reading each response before sending the next prompt.
2. Run Chain 2 against the real SkillBridge transcript the same way.
3. Then write Prompt 1 for the big task you brought from last lecture.

Before Lecture 9.3

Naming the Pattern

Last lecture, without naming it, you actually used the same specific, repeatable 3-step method twice, once on a campaign report, once on a meeting transcript. Today we're putting a name on it so you can reach for it deliberately on any big job from now on: Extract, Connect, Conclude.

Step 1, Extract

Pull out the raw information cleanly, with no interpretation attached yet. This step is about accuracy, not insight.

Step 2, Connect

Look for relationships between the pieces you just extracted. This step is about pattern-finding, still without jumping to a final answer.

Step 3, Conclude

Now, and only now, ask for the final answer, decision, or recommendation, built entirely on the extraction and connections from the previous two steps.

Try It Yourself

1. Pick a new big task, different from the ones we've used so far.
2. Write all 3 prompts in the Extract, Connect, Conclude pattern before running any of them.
3. Run them in order, reading each response before sending the next.

Before Lecture 9.4

Not Every Task Needs This

Task decomposition is powerful, but using it on every single prompt would slow you down for no reason. Today we're learning to tell the difference between a job that genuinely needs to be broken apart and one where a single strong CRAFT prompt is already enough.

When One Prompt Is Enough

• The source material is short enough to reason about all at once (a page or two, not 40)
• There's one clear deliverable, not several different angles layered together
• You already know roughly what you're looking for

When You Need to Decompose

• The source is long or complex, a full report, a large dataset, a lengthy transcript
• You're asking for several different kinds of analysis at once (extract, then interpret, then recommend)
• You genuinely don't know what you're looking for yet, and need the first response to guide your next question

Example 1: A Customer Complaint

Example 2: The AuraTech Report

Example 3: The SkillBridge Transcript

In each pair, the first question has one short, factual answer, zero decomposition needed. The second question layers several different kinds of analysis on top of a long source, exactly the kind of job that overwhelms a single prompt and needs Extract, Connect, Conclude instead.

Try It Yourself

1. Take the small task and the big task you brought today.
2. Run the small one as one CRAFT prompt, complete in one shot.
3. Sketch out how the big one would break into Extract, Connect, Conclude stages, you don't have to run it yet, just plan it.

Before Lecture 9.5

Full Decomposition, Start to Finish

This is the hands-on close to Module 9. No new theory today, we're taking real documents all the way through Extract, Connect, Conclude, using everything from this module together, guardrailed with constraints from Module 7.

The Full Process

• Step 1, confirm this task actually needs decomposing, using Lecture 9.4's checklist.
• Step 2, write your Extract prompt, raw information only, no interpretation.
• Step 3, run it, and actually read the response closely.
• Step 4, write your Connect prompt, referencing the response, not repeating raw data.
• Step 5, run it, read it, then write your Conclude prompt asking for a clear, committed final answer, guardrailed with a constraint.

Walkthrough 1: The AuraTech Report, Guardrailed

Walkthrough 2: The SkillBridge Transcript, Guardrailed

Walkthrough 3: Your Own Task

Now do the same thing with the real task you brought today, run the full 5-step process, and add your own guardrail constraint on the final Conclude prompt.

Your Turn

1. Run Walkthrough 1 against the real AuraTech report.
2. Run Walkthrough 2 against the real SkillBridge transcript.
3. Run the full process on your own task from Walkthrough 3.
4. Read each final Conclude response. Does it feel more trustworthy than if you'd asked for everything in one giant prompt?

Module 9 Recap

• Task decomposition means breaking a big job into a sequence of prompts, each building on the last.
• Extract, Connect, Conclude is a reusable 3-step pattern for almost any complex analysis, reports or transcripts alike.
• Not every task needs this, short, single-purpose tasks are still a job for one strong CRAFT prompt.
• Decomposition and CRAFT combine well, each link in the chain can still use a role, format, and constraints.

You can now take on jobs that would have felt impossible to prompt for at the start of this course, and break them into something completely manageable.

Before Lecture 10.1

You Didn't Learn 9 Tricks. You Learned One Skill.

We're at the last module of Part 1, and I want to reframe everything you've done so far. It can feel like you learned 9 separate techniques: zero-shot, role, few-shot, structure, chain-of-thought, constraints, CRAFT, and decomposition. Really, you learned one skill, giving an AI direction, expressed 9 different ways. Today we start stacking them together the way real, professional prompts actually get built.

Example 1: Diagnosing Customer Churn

Example 2: A Product Launch Email

Example 3: A Security Code Review

Count the techniques in each one, role, context, few-shot, chain-of-thought, format, constraints, five or six techniques, one prompt, every time. This isn't more complicated than what you've already written, it's just everything you know, organized in one place.

Try It Yourself

1. Take any task from earlier in this course.
2. Try adding just one more technique on top of what you already used for it, a role if you didn't have one, examples if you didn't have any, constraints if you had none.
3. Notice: stacking rarely means starting over, it usually means adding one more labeled block.

Before Lecture 10.2

Case Study: Three Full Marketing Workflows

Today we're walking through three complete, real-world workflows from start to finish. Watch how every module in this course shows up somewhere in each sequence, no matter how different the businesses are.

Case Study 1: A Bakery Win-Back Campaign

Case Study 2: A Gym Membership Win-Back

Case Study 3: A Nonprofit Donor Re-Engagement

Same shape, three completely different organizations. That's role and context and format and constraints in Prompt 1 (Modules 3, 5, 7, 8), few-shot in Prompt 2 (Module 4), and chain-of-thought in Prompt 3 (Module 6). Nothing here is a new idea, it's the whole course working together, every time.

Try It Yourself

1. Take the marketing task you brought today.
2. Write your own Prompt 1 using full CRAFT.
3. Write Prompt 2 adding real few-shot examples of your own voice.
4. Write Prompt 3 as a chain-of-thought gut check before you'd actually use the output.

Before Lecture 10.3

Case Study: Three Full Technical Reviews

Marketing isn't the only place this stacking approach shines. Today we're running the same combined approach through three completely different technical domains, using everything from the CRAFT lectures back in Module 8, plus decomposition and chain-of-thought.

Case Study 1: A Data Pipeline Review

Case Study 2: A Login System Security Review

Case Study 3: A Mobile App Performance Review

All three mirror the exact same structure, CRAFT foundation first, then a deeper reasoning pass, then a focused follow-up, even though the subject matter ranges from data pipelines to security to mobile performance. That's the real proof this is one transferable skill, not 9 unrelated ones.

Try It Yourself

1. Take the technical task you brought today, or use one of the 3 case studies in a new domain.
2. Write your own 3-prompt sequence: CRAFT foundation, chain-of-thought deep dive, focused follow-up.
3. Run all three and evaluate: does the final output feel like it came from a genuine expert reviewer?

Before Lecture 10.4

Building Your Prompt Library for Life

Over the last 48 lectures, you've saved roles, few-shot examples, structured templates, CRAFT prompts, and full prompt chains. Today we're organizing everything into one lasting system, so this course keeps paying off long after it ends.

How to Organize Your Library

• By task, not by technique. Organize around what you actually do ("client emails," "weekly reports," "code reviews"), not around lecture numbers.
• One finished template per task, already stacked with whatever roles, examples, format, and constraints it needs. You shouldn't have to rebuild it from parts each time.
• Clear labels, exactly like we labeled few-shot examples back in Module 4, so future-you can find the right one in seconds.
• A short changelog, if you improve a template, note what changed and why, so you don't accidentally undo a good fix later.

Three Sample Library Entries

Try It Yourself

1. Go through everything you've saved across this entire course.
2. Sort it by task, not by module or technique.
3. Pick your 5 most useful entries and rewrite them in the sample format above.
4. Delete or merge anything outdated or duplicated.

Before Lecture 10.5

Capstone Presentations and Part 1 Wrap-Up

This is it, the last lecture of Part 1. Today is about two things: presenting the work you've built, and stepping back to see how far you've actually come since Lecture 1.1.

Capstone Presentation Format

• Present one real task or project you solved using this course.
• Show the before, your very first, roughest attempt at prompting for it.
• Show the after, your final, fully stacked prompt or prompt chain.
• Name every technique you can spot in your final version: role, few-shot, structure, chain-of-thought, constraints, CRAFT, decomposition.

Look How Far You've Come

In Lecture 1.2, we typed "best gym marketing ideas" into an AI tool and got something generic. Now, in Lecture 10.5, you can write a 6-technique stacked prompt, or break an entire report into a 3-stage reasoning chain, without even thinking about it as separate steps anymore. That shift, from typing search-engine keywords to genuinely directing an AI, is the entire point of this course.

Part 1 Full Recap

• Module 1, Foundations: AI follows instructions, it doesn't search. Every prompt is built from 5 blocks.
• Module 2, Zero-Shot: the baseline, great for simple tasks, limited for anything nuanced.
• Module 3, Role: a lens, not just a title, changes tone and rigor.
• Module 4, Few-Shot: showing beats telling, especially for voice.
• Module 5, Structure: define the shape, not just the content.
• Module 6, Chain-of-Thought: reasoning beats guessing on hard problems.
• Module 7, Constraints: guardrails aim your prompt, they don't weaken it.
• Module 8, CRAFT: one checklist that bundles everything together.
• Module 9, Decomposition: break big jobs into a chain of prompts.
• Module 10, Capstone: real prompts stack several of these at once.

Part 1 is complete. Everything from here forward in Part 2 is applying this exact same skill set to specific, real content types you actually create.

Before Lecture 11.1

Welcome to Part 2

Part 1 gave you a complete, transferable skill: giving an AI clear direction. Starting today, we're applying that exact skill to specific content types, beginning with one of the most exciting and fastest-moving areas in AI right now, video generation. Whether you're using tools like Runway, Sora, Kling, Luma, or Higgsfield, the prompting principles in this module apply across all of them.

Why Video Prompting Needs Its Own Rules

Everything in Part 1 assumed a back-and-forth conversation, you write a prompt, read a response, refine, repeat. Video generation tools work differently. There's usually no reasoning happening at all, no role, no chain-of-thought, just one dense prompt that gets translated directly into a few seconds of visual motion. That means precision and visual vocabulary matter more here than almost anywhere else in this course.

Example 1: A Beach Walk

Example 2: A City Street

Example 3: A Mountain Hike

In every case, the weak prompt leaves every visual decision to chance: what does the scene look like, what time of day, what kind of shot. The strong prompt answers those questions before generation even starts, which is exactly why it produces a far more intentional, usable result.

What's Different About This Module

You won't be using role prompting or chain-of-thought the way you did in Part 1. Instead, over the next 5 lectures, you'll learn a new vocabulary: how to describe subject, action, camera, lighting, and style with enough precision that the model has almost nothing left to guess.

Try It Yourself

1. Think of one short scene you'd like to generate, real or imagined.
2. Write it as a single, plain sentence, the way you might describe it casually.
3. Keep this sentence. We're going to rebuild it properly across this entire module.

Before Lecture 11.2

The Five Parts of a Video Prompt

Just like CRAFT gave you five blocks for a text prompt, video prompting has its own five blocks. Once you can identify these five in any prompt, writing a strong one becomes a fill-in-the-blanks exercise, not a guessing game.

• Subject, who or what is in the scene, described physically
• Action, exactly what's happening or moving
• Camera, the shot type and any camera movement
• Lighting, time of day, mood, light quality
• Style, the visual aesthetic, cinematic, animated, documentary, film stock

Example 1: A Man Feeding Pigeons

Example 2: A Chef Plating a Dish

Example 3: A Dog Running on a Beach

Notice all three read almost like a shot description from an actual film script. That's not a coincidence, video prompting borrows directly from cinematography language, which is exactly what we're digging into next lecture.

Why All Five Matter

Leave out Lighting, and the model picks an arbitrary time of day. Leave out Camera, and you get a random, often oddly-framed shot. Every blank you don't fill in gets filled in by the model's own guess, exactly like we saw with zero-shot text prompting back in Module 2.

Try It Yourself

1. Take the one-sentence scene idea you brought today.
2. Break it into the 5 labeled parts: Subject, Action, Camera, Lighting, Style.
3. Fill in any part that was missing from your original sentence.
4. If you have access to a video tool, run it and see how much more intentional the result feels.

Before Lecture 11.3

Speaking the Camera's Language

Of all five video prompt blocks, Camera is the one beginners skip most often, and it's often the highest-leverage one to get right. Today we're learning the actual vocabulary filmmakers use, so you can describe exactly the shot you want instead of hoping for one.

Shot Types

• Wide shot, shows the full scene and surroundings
• Medium shot, from roughly the waist up, good for conversation or clear action
• Close-up, focused tightly on a face or object, high emotional detail
• Over-the-shoulder, framed from behind one subject, looking at another

Camera Angles

• Eye level, neutral, natural feeling
• Low angle, looking up, makes a subject feel powerful or large
• High angle, looking down, makes a subject feel small or vulnerable

Camera Movement

• Static, no movement at all, often the most reliable choice
• Pan, camera pivots left or right from a fixed point
• Tracking shot, camera moves alongside the subject
• Dolly in / dolly out, camera moves toward or away from the subject
• Drone shot, an aerial view, often sweeping

Example 1: A Car on a Highway

Example 2: A Market Walk

Example 3: A Bird Over a Forest

Across all three, naming the shot type, angle, and movement gives the model almost nothing left to guess. The vague version leaves the actual framing entirely up to chance.

Try It Yourself

1. Take your 5-part prompt from last lecture.
2. Rewrite the Camera block using specific shot type, angle, and movement vocabulary from today.
3. If possible, generate both the vague and precise camera versions and compare.

Before Lecture 11.4

The Consistency Problem

Here's a real limitation you need to know about: AI video models generate each clip somewhat independently, which means the same character can look noticeably different from one clip to the next unless you actively fight against it. Today we're learning the techniques that keep characters, scenes, and style consistent across multiple generated clips.

Technique 1, The Reusable Character Sheet

Write one extremely detailed physical description of your character once, then paste that exact same wording into every prompt that includes them. Consistency comes from repeating identical language, not from re-describing the character slightly differently each time.

The same technique works whether you're locking down a person, a mascot, or an entire recurring location.

Technique 2, Lock Your Style Language

Just like the character description, reuse the exact same Style and Lighting wording across every clip in the same sequence. If clip 1 says "cinematic, muted color palette, 35mm film look," every other clip in that story should say the exact same thing.

Technique 3, Use Image-to-Video When Available

Many tools let you start from a reference image instead of text alone. If your tool supports this, generate or upload one strong reference image of your character or scene first, then use that same image to anchor multiple clips. This is currently the most reliable consistency method available.

Technique 4, Keep Scenes Simple and Similar

The more different two clips are in setting, lighting, and action, the harder it is for a viewer to accept them as the same continuous story, even if the model does a perfect job. Keep transitions logical: same time of day, same general location, unless the story specifically calls for a change.

Try It Yourself

1. Take the multi-clip scene idea you brought today.
2. Write one detailed character sheet you'll reuse word for word.
3. Write 2 prompts for 2 different moments in the story, using the identical character and style language in both.

Before Lecture 11.5

When Video Generation Goes Wrong

Every AI video tool, no matter how advanced, has known weak spots. Today we're covering the most common failures, why they happen, and the specific prompting fixes that reduce them.

Failure 1, Warped Hands and Faces

Hands and faces are still genuinely hard for these models. Fix: avoid close-up shots of detailed hand movement (typing, playing an instrument, intricate gestures) unless your specific tool is known to handle it well. Favor wider shots or simpler poses instead.

Failure 2, Morphing or Melting Objects

Objects that shift shape mid-clip usually come from overly complex or physically unlikely scenes. Fix: describe fewer moving elements per clip, and avoid asking for physically unusual combinations.

Failure 3, Too Many Actions at Once

Most clips are only a few seconds long. Cramming several separate actions into one clip forces the model to choose which one to prioritize, and the result is often a confused blend of all of them. One clear action per clip is far more reliable, whether it's a kitchen or a street scene.

Failure 4, Garbled Text in the Scene

Signs, labels, and readable text inside a generated video are still very unreliable. Fix: avoid requesting visible text in the scene entirely, or plan to add text afterward in editing instead.

Failure 5, Ignored Camera Instructions

If the camera movement you asked for didn't happen, the request was likely too complex or combined with too much else. Fix: simplify to one clear camera movement, and make sure it's stated early and clearly in the prompt.

Try It Yourself

1. Look at the strange output you brought today.
2. Diagnose which failure type it matches from today's list.
3. Rewrite the original prompt applying the matching fix.
4. If possible, regenerate and compare.

Before Lecture 11.6

Storyboard Your 30-Second Film

This is the hands-on close to Module 11. Since most AI video clips run only a few seconds, telling a complete short story means planning multiple clips in advance, exactly like a real filmmaker storyboards a scene before shooting it. Today, you're doing exactly that.

The Storyboarding Process

• Step 1, write a one-sentence summary of your 30-second story.
• Step 2, break it into 4 to 6 short clips, each with one clear action (Lecture 11.5).
• Step 3, write a reusable character sheet if a character appears in more than one clip (Lecture 11.4).
• Step 4, write each clip's full 5-part prompt: Subject, Action, Camera, Lighting, Style (Lecture 11.2).
• Step 5, keep Lighting and Style language identical across every clip.

Storyboard 1: The Fisherman

Storyboard 2: The Coffee Shop Morning

Storyboard 3: The Kite Flyer

Notice the character's description and the Lighting and Style language repeat word for word within each story. That repetition is not accidental, it's the entire consistency technique from Lecture 11.4, applied across three completely different stories.

Your Turn

1. Take your 30-second story idea from today.
2. Run through all 5 steps of the storyboarding process.
3. Write out the complete prompt for every clip.
4. If you have access to a video tool, generate all the clips and watch them in sequence.

Module 11 Recap

• Video prompting is dense and descriptive, not conversational, there's usually no back-and-forth reasoning involved.
• Every strong video prompt fills in 5 blocks: Subject, Action, Camera, Lighting, Style.
• Camera vocabulary, shot type, angle, and movement, gives you real, specific control.
• Consistency across clips comes from reusing identical character and style language, not re-describing things slightly differently.
• Most failures come from overloaded prompts, simplify to one clear action per clip.

You can now plan and prompt an entire short AI film, not just a single lucky clip.

Before Lecture 12.1