Oct 6, 2025 · 50m · startup-ideas
I Watched Dan Koe Break Down His AI Workflow OMG
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth interview, host Greg Isenberg and creator Dan Koe break down Koe's end-to-end AI workflow, detailing how to combine large language models, research synthesis, and multi-platform repurposing to build a high-output media business. The discussion offers actionable frameworks for deconstructing viral mechanics, building two-phase meta-prompts, and translating digital audience attention into scalable startup ventures.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 21.8% of the talking time here. How this is scored →
Greg as informed peer 2.5 Guest teaching 6.5 Guest disagreement 0.2 Greg pushing back 0.6
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:40 Dan rejects relying on visuals over idea density
Dan dismisses the necessity of visual assets on social platforms, arguing that constraints on pure writing and high idea density build a stronger, more lasting connection with audiences.
Hardest push from Greg ▶ 26:07 Greg presses Dan on ignoring visual mediaGreg explicitly challenges Dan's text-only approach, citing a 40% performance lift with image posts and demanding to know why Dan omits visual assets from his playbook.
Biggest teaching moment ▶ 35:10 Dan breaks down viral post reverse-engineeringDan systematically shows how feeding top-performing posts into LLMs extracts psychological blueprints that can replicate high performance across any niche.
Greg holds their own ▶ 29:45 Greg demonstrates the power of visual media using his own tweetGreg navigates directly to his live account to demonstrate with real analytics how adding an image turned a flatlined tweet into 148,000 impressions.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Setting the Stage: Demystifying Dan Koe's Output | 2 | 5 | 0 | 0 | Greg introduces Dan and sets the expectation for Dan to share his full AI workflow without holding back. Dan outlines his high-level content ecosystem centered on newsletters and X posts. | |
| Mid-Roll Sponsor: IdeaBrowser.com | 1 | 5 | 0 | 0 | Following an ad read, Dan explains how he uses LLMs with large context windows to synthesize multi-hour video transcripts into condensed research notes. | |
| Formulating Core Social Ideas and Swipe Files | 4 | 6 | 0 | 1 | Greg summarizes Dan's system and presses him on the initial ideation process. Dan educates Greg on maintaining an active swipe file and synthesizing concepts under a personal brand. | |
| Deconstructing Tweet Mechanics: Idea vs. Structure | 2 | 7 | 0 | 0 | Dan demonstrates how to separate tweet structure from the underlying concept, showing tools like SuperX to deconstruct and remix successful formats. | |
| Repurposing to YouTube: Title Generation and Scripting | 1 | 7 | 0 | 0 | Dan shows his specific YouTube title generation prompt and addresses the common objection of cross-posting identical concepts across multiple platforms. | |
| Deep Post Prompting: Extracting Psychological Building Blocks | 1 | 8 | 0 | 0 | Dan breaks down his deep post prompt that deconstructs newsletters into psychological components like paradoxes, core problems, and transformation arcs. | |
| Daily Creative Routine and the 70/30 Follower Engine | 2 | 7 | 0 | 0 | Dan outlines his 70/30 follower acquisition formula, balancing proven spin-offs that reliably attract followers with regular experimentation. | |
| Visual Media vs. Pure Idea Density | 6 | 4 | 2 | 5 | Greg challenges Dan's text-only model by presenting his own tweet data proving visual assets boosted performance, while Dan defends text constraints and idea density. | |
| The Reverse-Engineering Framework for High-Performing Content | 2 | 8 | 0 | 0 | Dan demonstrates a live walkthrough of reverse-engineering viral posts into architectural frameworks using structured Claude prompts. | |
| Building Two-Phase Meta-Prompts for Content Generation | 1 | 8 | 0 | 0 | Dan walks through creating a two-phase meta-prompt that first interviews the creator for context before generating diverse post variations. | |
| Applying AI Frameworks to Business Offers and Scaling Accounts | 5 | 7 | 0 | 0 | Greg shares his holding company perspective on scaling content across teams. Dan expands the prompt framework to business offer creation and rapid iteration. |
Statements from this episode (16)
Dan Koe: Content Workflow Relies on Claude and ChatGPT
“I'm not a coder. I don't use too many of the fancy tools, mostly just cloud chat GPT.”
Dan Koe: Recreate top niche topics without watching the videos to stand out
“So a good way to generate ideas that help you grow is to go to YouTube, look at the accounts in your niche, Or the people, the kind of like tribe you want to join online. Go to their account, go to videos, filter by most popular, and then just write down 10 of…”
Koe: Does not use LLMs to write his actual newsletters
“Now I don't have the LLM right for me, like the actual newsletter.”
Koe: Uses Gemini For Three-To-Six-Hour Video Summarization
“I tend to go with Gemini 2.5, especially when there's, oh, this one's only nine minutes, but for like the three to six hour videos, I'll use Gemini just because the large context window.”
Dan Koe: Any idea can succeed using proven tweet structures
“Really any idea has the potential to do very well. Cause you can just pop it into a preexisting structure, or you can take a structure and plug in one of the ideas that you're trying to include in your brand.”
Koe: Repurposing Content Across Platforms Does Not Bore Audiences
“The objection there usually is, ah, well, if you're posting the same thing on two platforms, won't people notice that, or won't they get bored? Right. And most of the time, I've seen the opposite. Like, there are people that read it, and then when they watch o…”
Koe: Better to Produce One Great Weekly Asset Than Platform-Specific Mediocrity
“My philosophy behind it is I'd rather produce one amazing thing a week and just put all of my attention into that And put it out across all platforms rather than try to create something new for each platform and each of those things not be as good.”
Koe: LLMs struggle to write tweets without hyper-specific guidance
“So with this, I'm not asking it to write tweets for me because the LLM isn't going to be too good at doing that unless you
Guide it to in a very specific way, but instead here is I have it give me the pieces of ideas like the deconstruction of high performing …”
Dan Koe: Daily two-hour writing sessions power my entire content output
“How I actually structure that on a day-to-day basis is quite simple, where I wake up, do a little morning routine, shower, walk, whatever it may be, and then I sit down at my computer and the next two hours is just dedicated to writing. So in that writing, I h…”
Dan Koe: Balancing core proven ideas with experimentation drives all social platform growth
“And if you keep that balance of your core ideas that are consistently going to bring in followers with a healthy dose of experimentation until you find more, And then you just cycle out the core ideas. That's really the entire growth process on any social plat…”
Isenberg: Pairing text with images boosts X post performance by 40%
“So on X, I've noticed that like if I tweet just words versus I tweet words with an image, the words with image
You know, call it performs 40% better.”
Isenberg: Follower counts do not matter and smaller accounts get easier impressions
“The followers don't really matter anymore. Realistically, you know, it's actually easier to get impressions if you have less followers.”
Koe: Reverse-engineers high-performing content into LLM meta-prompts
“Anything that I like and want to incorporate into my work, I will ask Claude or ChatGPT to break it down, teach me how to do it. I will take that as if it were a, an SOP or just a set of instructions. Then I will take a prompt that helps me create prompts, sen…”
Koe: Two-phase prompt framework requires AI to interview the creator first
“Structure this prompt in two phases. The first is context gathering, so interview me for the ideas that I want to write about, and phase two is post writing, so write three variations of each type of post using the ideas I gave you.”
Koe: Prompting LLMs with specific expert frameworks beats general prompts
“The reason we do this rather than, hey, here's my product. Help me create an offer from it is because I want to narrow the context from the beginning, right? I don't want AI. I don't want the LLM to pull just general context that it thinks is best for creating…”
Koe: Two-phase prompting matches 2-3 months of marketing study
“It spits out an offer blueprint that is probably 10 times better than something you would have done on your own just then, right? And it would probably be on par with what you would be able to do after Two to three months of actually learning about it.”