Welcome to Week 4
Advanced Prompt
Engineering
Logistics
● Don’t forget about the context engineering challenge!! $1500 to first!
● Tuesday at 12 pm et - guest session - deep dive into Poppy AI
● Week 5: CustomGPTs + Poppy AI
● And now, we look at advanced prompting
The 8-Week Program
You are here - last week
of fundamentals!
Prompting &
Context
Automations Workflows
BONUS
Sessions
Quick Recap
● Week 1: The future of everything is going to be prompting
● Week 2: We covered prompting fundamentals
● Week 3: We covered context engineering
● And now, we look at advanced prompting
Today’s agenda
● How to use/prompt Deep Research
● Intro to Operator & Manus
● How to prompt Image Gen
● How to prompt Veo 3
So far, we have looked at the strategist
We have primarily
focused here
Now, we are going to look at the actual generation
We are going to cover
generation today!
We are covering only one session on ad generation
because honestly, it’s not great yet.
But nonetheless, it’s important to know how to prompt
these tools!
Deep Research
How we use it
● Filling out context libraries (extremely valuable)
● Come up with headlines
● Finding insider jokes/phrases
● Creating new customer personas
How to prompt it
Task
T h o u ght
of
Chain arch)
(rese
T h o u ght
of
Chain l draft)
(initia
of
Chain
ht
Thoug ut)
outp
(final
Prompting Hack: GPT 4.5
FYI, Claude also has a research function!
Manus // ChatGPT Operator
I’m not going to spend a lot of time covering these
tools because honestly, they are not great.
BUT, they will be.
So far…
1. Use Reddit answers, deep research, Manus/operator, good prompting
+ context to help think of good ad ideas (research)
2. Now, we will move onto Image Gen
Image Gen
How we use it
● Product image generation
● Static ad generation
● Landing page inspo
How to prompt it
1. Get an example of the output you need from Claude
2. Go to Claude and upload your product image + reference design
3. Go to ChatGPT and paste the overview you got from Claude
Step 1
Step 2
Step 3
Initial Output
Overall, not bad
Step 4
V2 Output
Essentially you can keep
asking it over and over “how
can we make this better?”
V3 Output
But honestly, I feel like it kinda
gets stuck at this 80% range.
V4 Output
Again, it’s decent.
You can select specific
One other feature to note areas to edit here
One other way
Output
Other Use Cases for Image Gen
Unrealistic Visuals
Prompt:
A close-up photo of a woman’s bare midriff
with the text “NO MORE FATIGUE, HOT
SWEATS OR WEIGHT GAIN” spray-painted in
bold black letters across her stomach. She is
wearing a cropped tank top and sweatpants.
She is holding the product image I am
attaching in her hand near her waist. Natural
indoor lighting, realistic photography style.
Product Variants
Prompt:
"Take this ad image and update it to match
the [new variation]. Change the packaging,
background, and supporting visuals to
reflect the [new variation] while keeping the
layout, model, and text style the same."
So far…
1. Use Reddit answers, deep research, Manus/operator, good prompting
+ context to help think of good ad ideas (research)
2. Image gen can help us generate statics
3. Now, we are on to video!
Veo 3
Video Generation
1. Arcads is a great platform & you all got the demo of that!
2. Runway ML, Sora, and Veo 3 are the best text-to-video models
3. BUT, Veo 3 is the best today
Examples
● Watch this video
How to prompt it
1. Cinematic hook
2. Subject description
3. Action and movement
4. Setting and environment
5. Lighting and color
6. Camera movement
Putting it all together
Task: Create a video of a guy crying out of frustration
Honestly, prompting Veo 3 is tough.
1. Google’s guide to prompting Veo
2. Dave Clark’s Tweet
3. Curious Refuge
So, I just took a lot of these guides and created a custom
Google AI studio project :)
Here is my process
Step 1
Step 2
Step 3: https://labs.google/flow/about
Step 4
Step 5
Knowing what to prompt is what we will focus on
1. Full AI commercials (very hard)
2. Single shot clips (4-6 seconds, more manageable)
How we use it
● B-roll clips of hyper-unrealistic shots
● Realistic content that would be extremely hard to shoot
● Testing full commercials
You need to get CREATIVE
1. Anacondas coiled around a man’s legs
2. A fake “Shark tank” pitch that says something vulgar/shocking
3. A lot of intense emotions - crying, anger, humor
4. Police chase helicopter POV ending with a guy wearing TPJ and everyone laughing
5. First ever footage of a man with “noasssyndrome”
6. Man in the Oval Office, selfie style, saying “So I’m chilling here in the oval office -
did you guys know they have a framed pair of jeans?”
WARNING: you can waste a LOT of time doing this…
IMO: Unless you have extensive production experience,
there is a pretty big learning curve.
This is going to fundamentally change advertising. Like
we mentioned in week one, EVERYTHING will be a prompt.
Exercise one
1. Create one static ad good enough to where your creative team would
be willing to test it!!
Exercise two
1. Spend two hours playing around with Veo 3 and coming up with
prompts.
Next Week: Building Poppy AI +
CustomGPTs!
