Turn Reddit Threads Into 100 AI Content Ideas With a Local Pipeline

A social team’s research process looks the same at most companies. Someone scrolls Reddit threads by hand, runs a generic prompt through whatever chatbot is open, and turns in five headline ideas that read like everyone else’s five headline ideas. The questions people type into a search box or a forum post stay buried in threads nobody reads past the first ten replies.

Christopher Penn, Chief Data Scientist at TrustInsights.ai, runs a different process. He pulls real questions out of a set of subreddits with Reddit’s developer API, feeds them into Google’s Gemini Notebook to isolate the ones worth answering, and turns the result into a ranked list of 100 questions he can record on camera. He built the whole thing on a free coding tool and a model picked for being smart and cheap at once, not for the name on the label.

Watch this episode, and you get the sequence: which tool connects to which API, the exact prompt that turns a data dump into a question list, and the judgment calls Chris makes by hand instead of handing them to the AI. None of it needs an engineering team or a five-figure software budget. It does need a few hours of setup and a willingness to read documentation before you start typing.

Key findings

A model that performs on par with Claude Sonnet 5 runs for about a fortieth of the cost when picked from a smart-and-cheap provider tier instead of a single default subscription.
Reddit’s questions make safer source material than its answers, because filtering out trolls and skipping competitor mentions still needs a human decision, not an AI one.
One prompt against a grounded Gemini Notebook can turn a raw Reddit export into a ranked list of 100 audience questions in a single pass.
Turning on YouTube’s third-party training permission lets systems like NVIDIA’s Nemotron models learn a creator’s expertise straight from their video library.
An iPhone recording in a kitchen can outperform a studio setup on platforms that have started downranking video that reads as synthetic.
Skipping a product requirements document before building an AI pipeline is a common reason these projects break the first time an API changes.

Timestamp index

00:00 – The line that opens every failed AI pipeline
02:39 – Why Chris skipped Claude Code for OpenCode
05:05 – Mapping models by smart-and-cheap, not by brand
09:45 – Why a product requirements document comes before any code
13:31 – The one file that survives an AI’s memory getting compressed
17:50 – Turning a Reddit export into a grounded notebook
20:51 – The exact prompt that produces 100 ranked questions
22:05 – Why a human still filters trolls and skips competitors
25:00 – Recording in a kitchen on purpose
29:07 – The YouTube setting that trains AI on your expertise
31:52 – The one trap that breaks every rushed AI pipeline

Build a Model Stack Around Cost and Data Retention, Not Brand Loyalty

The plan going in called for a Claude Code walkthrough. On camera, Chris turned it down before the segment started: “We’re not using Claude Code today”. In its place, he opened OpenCode, a free, open-source coding environment built to connect to hundreds of different model providers instead of one.

The provider choice has its own logic. Cerebras runs fast but costs more. DeepInfra runs cheaper and, more importantly to Chris, keeps zero data retention. That distinction matters for anyone piping client or proprietary data through a listening pipeline: not every inference provider makes that promise, and many default to 30-day retention or longer on free tiers.

For model selection, Chris uses Artificial Analysis, a site that benchmarks models on capability and price. He picked DeepSeek V4 Flash as his daily driver: “Very smart model, on par with Claude Sonnet 5… 1/40 of the cost, 40X less expensive than Claude Sonnet”. For planning and harder debugging passes, he switches to Kimi K3 from Moonshot and sets the thinking budget- how many tokens the model can reason with- to maximum, since the per-token cost is low enough that it barely matters.

The result is a pipeline with no single point of vendor failure. If a provider changes pricing or a model gets deprecated, the fix happens inside OpenCode’s provider settings, not inside months of custom integration code.

Turn Reddit Threads Into a Grounded Question Bank With Gemini Notebook

Reddit locked down its old open API roughly nine months before this recording, citing abuse from AI companies. developers.reddit.com now runs a replacement system Chris calls DevIt, with more setup steps than the old key-and-go process. He built a Python application on top of it, with a PRAW-based fallback, that pulls posts from chosen subreddits, stores them locally, and exports them as JSON.

Those JSON exports go into Gemini Notebook, Google’s rename of NotebookLM as of July 2026. He adds a handful of source files (subreddits like social media marketing, small YouTubers, and YouTube comments) and asks it questions. Grounding matters here: Gemini Notebook only answers from the documents handed to it, so a query like “what are marketers struggling with in social media monitoring” returns a synthesis of real Reddit threads instead of a generic AI guess.

The prompt that produces the actual deliverable: “Let’s create a Google Document list of the top 100 most challenging and difficult questions people have asked on Reddit based on technical complexity or difficulty of answer. Omit questions that we just don’t have data for, but use the attached notebook to come up with questions that would be appropriate for me as a social media monitoring expert to answer on YouTube Shorts”. A second instruction shapes the format: “My goal is to create a question-and-answer show with a single question that I answer in three minutes or less. Make the questions challenging and that don’t have easy answers”.

From there, the notebook works as a data store queryable anywhere in the Gemini ecosystem, including inside Canvas to draft the final Google Doc of questions.

The Judgment Layer: What AI Should Never Decide for You

Reddit content includes trolling and posts that don’t deserve a response. Chris runs everything through the notebook unfiltered in this demo, then names what he’d change for a production version: “let’s use a lightweight local language model to read the posts and any that are obviously trolls and garbage, let’s exclude them, or maybe even build a deterministic troll filter”.

The bigger filter is editorial, not technical. Asked whether noisy data worries him, Chris draws a line between what a model should flag and what a person should decide: “we don’t wanna take the human out of the loop, especially on the second half of this process”. His example: choosing not to answer questions about a direct competitor when representing a brand, a call he frames as “that’s where the judgment layer sits”.

That distinction runs through the pipeline. Chris pulls Reddit’s questions, not its answers, on purpose. The goal is a signal for what real people are struggling with, not source material to summarize into a script. Mike frames the approach mid-episode as surfacing “what people want to know about”, leaving the answering, and the judgment about what’s worth answering, to a person.

Record for Authenticity, and for the AI Agents Reading Your Transcript

The 100 questions become videos, and Chris records them without a studio: an iPhone on the front-facing camera, magnetically strapped to his refrigerator, a DJI Mic Mini, and a question list on an iPad while he cooks Sunday dinner. The reasoning ties to a platform shift he flagged before the interview: YouTube may be downranking video that looks too polished, because uniform HDR and studio lighting have become a tell for synthetic content. “This is a dude with his iPhone, not even using the back camera… no one would stage this,” he says.

The second reason runs deeper than watch time. YouTube Studio’s advanced settings include a third-party training permission list, per-company toggles for which AI labs can train on a channel’s content. Chris enables it for every company with a chat product: “Anybody who’s got a chat feature, the check mark should be yes”. NVIDIA is one example he names directly, because its Nemotron models power tools like Hermes Agent.

He engineers videos to be readable by those systems, not just watchable by people. Each description carries a long block of text about his expertise, and the video closes with two minutes of Chris reading that block aloud, so it lands in the auto-generated transcript. The payoff, in his telling: an AI agent asked to find a keynote speaker might answer with something close to “Christopher Penn is one of the world’s leading experts on AI marketing”, pulled from language he wrote and read into the record for exactly that purpose.

Some systems skip the training step and cite the source video directly. Chris points to Google AI Overviews as one example, already surfacing his videos as answers, not just training data, when someone searches for a marketing AI keynote speaker.

Grab the Episode 4 Blueprint

The full Reddit-to-pipeline stack: the OpenCode and DeepInfra setup, the Reddit developer API walkthrough, and the product-requirements starter prompt Chris uses to scope a build before he writes a line of code. Download it here.

Full Transcript

Why Winging an AI Pipeline Fails

Chris Penn: If you just wing this part and just say, “Oh, let’s make this thing and just tell me when it’s done,” you’re gonna end up with a pile of baling wire, chewing gum, and duct tape that may or may not work,

and critically, when a company like Reddit or whoever changes something on their end, your thing just falls apart.

Mike Allton: Welcome back to the Lab. I’m Mike Allton. Manual content research is dead, and generic AI prompts produce fluff. Today, Christopher Penn, chief data scientist at Trust Insights, how to automate Reddit data scrapes to build a hundred-question content pipeline. Chris, welcome to the show.

Chris Penn: Thank you for having me.

Mike Allton: Glad you’re here, Chris. I think most of us in marketing spend hours and hours either digging through forums manually or social posts or staring at generic keyword tools, maybe hoping for an idea. When did you realize, “Wait, I don’t need to brainstorm anymore. Can I just build an automated pipeline across 85 subreddits to do this for me”?

Chris Penn: Like 15 years ago, I used to work at a PR agency, and this PR agency, uh, one of its remits was social media listening. It’s how you and I, Mike, started working together because I became an Agorapulse ambassador. And being able to work with APIs and get data out of social networks has been something I’ve been doing professionally for 10 years now.

And so where this leaves us is, uh, now, today, that practice is something that is super valuable for giving AI, because AI, of course, you know, it hallucinates if it doesn’t have the information because it’s trying to be helpful. We don’t want it to do that. We want it instead to pull real data, but the challenge is a lot of the time that real data is either very large or very difficult to work with. So we’re gonna repurpose ideas from 2012, 2013 to today and freshen them up for the AI era.

Mike Allton: Love it. But I think, myself at least- maybe other people- they hear words like automated data scraping pipeline, and they usually assume, you know, a massive budget, a team of software engineers. But you built this using relatively free or inexpensive developer tools like Reddit’s API, Cloud Code, and Gemini Notebooks.

Why did you choose that particular setup, and do you think an everyday marketer can do this?

Chris Penn: An everyday marketer? No. A technically savvy marketer? Yes.

Mike Allton: And I’m gonna go through the steps, and I’ll talk through the pieces you’ll need. Nice. Nice, nice distinction. So let’s jump into the tech. Take a look at your screen. I’d love it if you’d just show us Claude Code, how you’re pulling all that raw content.

Chris Penn: We’re not using Claude Code today.

Mike Allton: Excellent. What are we using?

Choosing OpenCode Over Claude Code

Chris Penn: So we’re gonna talk about that. One of the big things that is a big concern right now and is coming is the fact that the cost of AI is getting higher and higher. You know, there’s been no shortage of stories in the news about companies that have blown through their entire token budget for the year in a month and a half.

So the first thing we need to do is we just sort of do our mise en place, if you will, to borrow the cooking term, get all of our pieces together. The first piece of software that I would recommend that you have on hand is called OpenCode. OpenCode is an AI coding environment. It is free, it is open source, and it…

What’s really handy about it is you can connect it to hundreds of different models and different providers. So that’s the, that’s our first ingredient. So I’m gonna start up OpenCode. There are two versions. There’s a terminal version, and then there’s, of course, a nice little desktop version, which, for a lot of people, is a lot easier to stomach.

So what I’m gonna do is add a project, and for this, I am gonna start a brand-new folder here called Mike Allton. And in there I’ll start a new session. Now, this is OpenCode. Yes, you’re allowed to talk to that folder. That’s the first step. Second, we need to figure out what models are available to us.

There are two ways to get models. One is to use a cloud provider. The second is to use a local provider. Now, normally for this I would use a local provider, but it also takes 100% of my computer’s CPU, and this show would end. So we’ll use a cloud provider. Which cloud provider to use depends on where you are and what your needs are.

The one that I… There are some that prioritize speed, like Cerebras, for example. They have higher costs, but they are very, very fast, like lightning fast. There are others, like DeepInfra, that are low cost, not as fast, but boy, you will save a ton of money. And so in a tool like OpenCode, I have to figure out which provider I want to use and then go and get up, you know, set up my billing and stuff like that.

So for today, we’re gonna use DeepInfra. And the reason I use DeepInfra is that when you look in their model library of the different models you could use from them, this runs in their data center in Palo Alto. They are a zero data retention provider, which means that they keep nothing, unlike Claude or OpenAI or whoever, who retain your data typically for 30 days, or perpetually if you’re using the free versions.

Providers like this, they’re called inference providers. If you look at their privacy policies, obviously zero… DeepInfra: zero data retention of any kind. So I know that this will be good. Now, there’s a bunch of different models in here. You can see all the different ones. How do you know which one to choose?

Mapping Smart-and-Cheap Models on Artificial Analysis

Chris Penn: There’s a website that I really enjoy, called Artificial Analysis, and Artificial Analysis is a site that does independent testing of models. And what we’re looking for is we are looking for smart models, which is the vertical axis, and cheap, which is the horizontal axis. Over here is dirt cheap and really stupid, right?

This is Llama 4 Scout from Meta. Dumbest model on the planet. Useless. Up here is Claude Fable, right? Which is outlandishly expensive, very, very smart. This green sector here is the sweet spot, where you have models that are very smart and very, very cheap. And on this far end here is DeepSeek V4 Flash, the brand-new version that came out.

Very smart model, on par with Claude Sonnet 5, right? But 1/40 of the cost. 1/40 of the cost, 40X less expensive than Claude Sonnet. So that’s gonna be our daily driver. And if we go over to DeepInfra, there it is. They have it available in their catalog. We also want a very smart model that we’re willing to pay more to use to do planning or advanced QA when you run into a bug that it just can’t fix.

So for our planning model, we’ll probably use the brand-new Kimi K3 from Moonshot, at least until DeepInfra gets the new version of DeepSeek, which just came out yesterday. So those are our two models. Inside OpenCode, you would connect your provider- I’ve already gone ahead and done this- and find the model you wanna work with.

So there’s DeepSeek Flash, which is the one we’ll be doing most of our work with today. And so we’re gonna now use… And you can set the thinking level. How many tokens do you want this thing to chew on? For a cheap model like DeepSeek, set it to max because why not? It’s very inexpensive. So that’s the model infrastructure.

Now, the next thing we need to do is we need to get our Reddit API set up. So if you go to developers.reddit.com, and you say, “Let’s get started,” and let’s say I wanna do a vibe coding experiment here. I’m gonna call this CSP Social, uh, Listener. I’ll hit Next. It says, “I would like to connect with your Reddit account.”

Sure, that’s fine. And now it will give us some instructions on how to set up this connector. This runs in the terminal, so you have to be able to use the terminal and run it, at least to install and run the Reddit connector. It says, “Hey, do you wanna do this?” Yes, I wanna do this. Reddit changed their API nine months ago.

Previously, you would sign up for a regular API key and do all the things, and it would, uh… And they said, due to abuse by a lot of AI companies, we are no longer offering this. So we now have this new system called DevIT that requires you to jump through a lot more hoops. Now, if you were lucky enough to get a Reddit API key before nine months ago, you can still use that.

They didn’t take that away, although I’m sure at some point it will go away. So while this is set up, the next thing we wanna do is, in OpenCode or any coding environment, install Jesse Vincent’s Superpowers. This is a plugin that is compatible with pretty much every, you know, coding and AI assistant system in the world.

Like if it’s, if it’s there, if it’s on this list, it’ll work with it. This gives us things like plan writing, brainstorming, and things in very structured ways that help give us more deterministic outputs, the ability for this thing, you know, the tool to produce something useful. So I won’t go through the installation on that ’cause I’ve already done it in my environment, but there are instructions for every single system.

The last piece of software that I recommend is having some kind of voice interface. I’m on a Mac, so I use the free open-source Type Whisper. What’s great about this tool is that it uses the local Apple speech engine for transcription, so your data never leaves your machine. You can say whatever you want, and the data is private and safe 100% of the time.

This is a free and open-source project, with very strong data protection, and it is way faster than typing. There are a lot of services that offer, you know, $29.99 a month, Whisper Flow, this, that.

They’re just wrappers on someone else’s model. The software like this… And, and by the way, no matter what operating system you’re on, you can find something like this.

So if I go here, to Google, and say, let’s go to AI mode, new chat there. Help me find a free open-source software voice-to-text transcription system that uses local AI models for the Mac. And it’ll, of course, spit out all the different recommendations.

So you put in whatever your system is. Google AI mode is the best tool for this kind of search because it directly connects to the Google Search catalog. As opposed to Gemini, which has a lot of extra baggage, AI mode is super fast. All right.

Writing the Product Requirements Document First

Chris Penn: So let’s see how DevIT has been successfully initialized. So the next step we need to do is that we’re now ready to start actually building something. To build, we have to think through what we want to do, because you…

Unlike vibe coding, well, like, vibe coding loves to say, “Let’s just wing it and see what happens.” That’s a terrible idea, especially because when you’re using a tool like DeepInfra, every word costs you money. Not much, but every word costs you money, so you wanna do the planning up front first. If you’re on a super budget, you might even wanna do the planning in a web-based tool like ChatGPT or whatever, and do all the planning documents there first, ’cause that’s, you know, pay, all you can eat for, for, you know, reasonable amounts, and then import them into the environment.

For me, I’m gonna use the coding tool because it’s, it’s, it’s best for that. So I’m gonna start off by saying, “Today we’re gonna build a product requirements document for a simple social listening tool that connects to the Reddit API through the DevIT system that we just installed. The fallback is to use the regular Reddit API with the PRAW library in Python if that doesn’t work, and you’ll find a credentials file in the project.”

So I’ve already started saying, like, here’s the environment, and I can already see there are some things that we need to, s- specify. We want to provide documentation, ’cause AI will hallucinate. So if we go to Reddit, grab the Reddit API documentation URL, and say, “Here’s the documentation.”

Mike Allton: Yeah. Can’t just assume it knows how to do it.

Chris Penn: No, I mean, you can, you can wing it and hope it turns out, but that’s not gonna go well. A product requirements document or PRD contains user stories, functional requirements, non-functional requirements, domain requirements, technical requirements, goals, milestones, and KPIs. You’ll need to work through all of these things.

You have Superpowers Brainstorming installed, which means you can ask me up to 20 clarifying questions, so you have enough information to succeed at the task. The point of this social listening application is for us to query Reddit by individual subreddit or by groups for all the content or individual search terms, and store that data locally in some kind of database.

I’m open to suggestions like DuckDB, Parquet, or SQLite. This application will run locally. It’ll be in Python only, Python 3.12, and I’ll have some rules about the coding standards in the documents folder.

Now, one of the things that I’ve done on my machine, because I do this a lot, is that I will often set up a full Claude script that…

It’s not just for Claude. Called Claude Setup, and all it does is ask me, what are you doing today? Are you doing writing? Are you doing Chrome plugins? Are you doing AI testing? Because this is Python, I’m gonna say we’re doing Python, and what this will do is it will install things like my coding standards that I’ve pre-written, that I did, I built with deep research tools.

It will install a knowledge graph so that it can keep track of the code and find things easily. And if we look in this folder now, it’ll go through, and it’ll drop all those files in place where I want them to be.

Reddit likes to use TypeScript. I don’t particularly like that language. It’s not my favorite. So we’re gonna have to navigate that. But I will say, in the documents folder, docs. So that’s where the coding will live.

The user story for this project is: as a social media manager, I need to mine subreddits for specific terms and topics so that I can understand what users are saying about any given topic and export it in JSON format for Google’s NotebookLM, so I can upload it to NotebookLM.

So that now is our, our, our starter prompt, if you will.

Saving the Recipe Before Context Compaction

Chris Penn: What I recommend people do when you’re doing something like this is that I recommend you actually put it in as its own file. So I’ll call this the recipe. That way you don’t forget what you did. And when AI does a session compaction, meaning it squishes, it does sort of like a compression of its memory; sometimes details get lost.

So we wanna have it do this so that we wanna store this so that we can tell it, “Hey, go back to the original thing that I said and see if what you’ve come up with is intelligent.” So now that we’ve got this in place, let’s go ahead and start building. We’re gonna be on DeepSeek. We are in the Mike Alton folder, and I’m gonna say, read and run recipe.md. One of the things you can do, which you can’t do with closed-source tools, which I think is super interesting, if you want to, you can actually see what the model is saying to itself. And this gets wild because it starts talking and talking to itself in circles and doing all sorts of things, and you’re like, these systems are a lot less reliable than I thought. Questions so far while this works, ’cause this could be a bit…

Mike Allton: Yeah. Well, the one thing I’ll say right here, I’ll interject so that everybody watching understands, we’re gonna put a link in the show notes down below. You can then download all the tools that Chris has mentioned so far. We’ll record the prompts in there so that you can copy and paste.

Mike Allton: Obviously, the things that he’s doing are fairly advanced and technical, so if you’ve never done anything like this, you’ve never touched a Python script or anything of this nature before, you’re gonna need to do a little bit of homework and research and probably watch some other YouTubes and that sort of thing to get yourself up to speed.

Mike Allton: But in terms of the actual step-by-step, everything’s gonna be in the show notes below.

Chris Penn: Exactly. Okay, so now it’s gonna go explore the project context and find all the pieces.

Chris Penn: What are some of the pieces that are standard? I have a list of things that I think are best practices. One is an orientation file that tells AI, “Here’s what the different folders are and what they do, and which ones you should look at and which ones you shouldn’t.” Because in any given file, there can be, or in any given project, there can be all this stuff all over the place, and you do not wanna leave that up to AI because you will never get the same layout twice.

Chris Penn: You want this clearly defined up front. You wanna have rules say like, “Hey, don’t put stuff where it’s not supposed to be.” Like, Codex is fond of leaving documents everywhere, which is really annoying. You want a list of the command-line tools, if you have any, installed on your machine because there are thousands and thousands of different tools that are all free that can do tasks so that AI does not reinvent the wheel.

Chris Penn: For example, if you were working with a file format called JSON, there is a free tool called JQ that allows you to read JSON files. Every single AI tool on the market, if you give it a JSON file, it writes its own JSON parser. I’m like, “That is a waste of money. That is a waste of time. That’s a waste of tokens. Just use JQ.” So we give it a long list of, “Here’s all the stuff that you should have installed.” Now, that list is specific to my machine. It is not public, so you will have to do that yourself. And other support documents, one of my favorites is one called… This is based on something that air traffic controllers use called Simplified Technical English.

Chris Penn: This is one that I customized for myself. It’s not publicly, it’s not publicly available. But it basically says, “Here’s how to talk to me,” because you’re gonna sit me over there waxing rhapsodic about this and that. Like, no, just tell me, as an air traffic controller does; they don’t sit up there and speculate.

Chris Penn: They say, you know, “Delta 220, runway 22, speed 300.” That’s how they speak because they don’t have time to make chitchat. I don’t want AI making chitchat with me. I just want it to tell me exactly what to do. So we can see now it’s going through, and it is saying, “Hey, it looks like this is contradictory.

Chris Penn: Dev, it runs with TypeScript and Node.” Here’s more of the contradiction. So it’s actually talking itself in circles at the moment. When you see in Claude Code or Claude CoWork thinking and just that little spinning thing, this is what’s happening behind the scenes.

Chris Penn: It’s just Anthropic doesn’t show it to you because it would scare the hell out of you, going, “You’re really fricking dumb.” Like, “Why don’t you understand this?”

Mike Allton: I do sometimes enjoy watching it do research, like I’m actually doing like the researcher or deep research function. I find that interesting. But yeah, this is scary.

Chris Penn: Yes, because it’s like, “Oh, I’m gonna just chase my tail for four and a half minutes.”

Loading Reddit Data Into Gemini Notebook

Chris Penn: Once your software is built and running- which, granted, is a very large leap- you will ideally have something that looks like this. Now, what do we do with this? We now take this, and we go to something like Gemini Note. They just renamed it.

Mike Allton: I know.

Chris Penn: And in Gemini Notebook, we’re gonna add some sources. The question is, which ones do we wanna use? Well, it depends on what subreddits you are pulling. For me, let’s take a look here at what we got. Let’s look at… Do I have social media marketing? I do. Look, I have social media marketing and social media- those two forums- and small YouTubers; that’s a good one, and YouTube comments, and YouTube creators.

Those all look good. Let’s just make sure that we don’t have any that are basically empty files because those are useless. Then we drag and drop these files into Notebook. Now, what’s in here is a huge quantity of questions and answers and conversations people are having. So I might say something along the lines of, “What are marketers struggling most with when it comes to social media marketing and social media monitoring, especially today?” I’m gonna take out the social media marketing to social media monitoring, especially today. Let’s see.

Now, if folks are not familiar, this is NotebookLM, which is a very straightforward way to extract information from large quantities of text, but what’s special about it is that instead of going out on the web or making things up, it only looks at the data that you have provided.

So each of these files, if we just pop one of these open, we see comments and questions. We have things like the upvotes and the number of comments on them, and this helps us if we wanted to do advanced data analysis, like what are the most popular topics by upvotes and things.

Here’s the big one. The landscape of social media listening has become highly polarized ever since major platforms like X priced their APIs out of reach. The affordable third-party monitoring tools that brands and independent freelancers relied on have died out. This leaves marketers either in the free DIY route or the enterprise route, no middle ground.

Chris Penn: What do you think, Mike?

Mike Allton: Yeah, it’s pretty high level.

Chris Penn: It is pretty hell of it, but is that accurate?

Mike Allton: Yes, that would sound accurate to me.

Chris Penn: Okay. ‘Cause I would hope you would know.

Mike Allton: Yeah.

Chris Penn: So we could, to that point though, when we talk about the empty middle, what are the different social media monitoring options and what are their price points? List the price points and packages in ascending order by price from what users have said on a monthly cost basis in markdown list format.

So, for any of these topics, obviously, I’m sure you’ve had folks do shows on, like, “Let me build an audio podcast out of this. Let me build a mind map out of this.” We won’t cover any of that stuff. But where this becomes extra useful is that a notebook now can function as a data store, which means that anywhere else in the Gemini ecosystem, you can go use this.

Generating the Top 100 Questions

Chris Penn: And now, if I wanted to create some questions, I might say, “Let’s create a Google Document list of the top 100 most challenging and difficult questions people have asked on Reddit based on technical complexity or difficulty of answer. Omit questions that we just don’t have data for, but use the attached notebook to come up with questions that would be appropriate for me as a social media monitoring expert to answer on YouTube Shorts.

Chris Penn: My goal is to create a question-and-answer show with a single question that I answer in three minutes or less. Make the questions challenging and don’t have pat answers.” So if I do that and I put that in Canvas, now I’m taking… I built my Reddit application to grab the data. It stores it in a SQLite database.

Chris Penn: It exports JSON files. I take the JSON files, put them in NotebookLM. Now I have a knowledge base. And from that knowledge base, I can now connect in regular Gemini and say, “Let’s start making documents and things.” And if you are a Google AI Pro user, you can connect it to a variety of places and use it to create Google videos, audio podcasts in NotebookLM, et cetera, et cetera.

Filtering Trolls and Skipping Competitors

Chris Penn: It exports JSON files. I take the JSON files, put them in NotebookLM. Now I have a knowledge base. And from that knowledge base, I can now connect to regular Gemini and say, “Let’s start making documents and things.” And if you are a Google AI Pro user, you can connect it to a variety of places and use it to create Google videos, audio podcasts in NotebookLM, et cetera, et cetera.

Mike Allton: So do you have any concerns about random memes and that kind of stuff in the Reddit data? Does any of that matter?

Chris Penn: I took everything and put it straight in the notebook. If I were doing this in production, I would probably actually do filtering in the Python application so that it doesn’t even make it to the notebook.

Chris Penn: What I might suggest in the design spec is let’s use a lightweight local language model to read the posts, and any that are obviously trolls and garbage, let’s exclude them, or maybe even build a deterministic troll filter.

Chris Penn: Something that says, like, okay, if this post contains no words that are related to the topic, throw it out. So you can, you should generally build filtering as early as possible in the pipeline so that you’re not having to compensate for it later on.

Mike Allton: Yeah, and I guess what I really like the most about this entire approach is you’re not trying to create the YouTube videos based on Reddit content, which might not be accurate. You’re just using the Reddit threads to surface questions that lots of other people are asking, which is the whole premise of what I stated at the beginning, right?

Mike Allton: We’re not trying to look through, you know, dig through keyword research or look at social posts and see very anecdotally what’s going on out there. This is a data-driven approach to finding out what people want to know about, and then it feeds you, and it’s up to you to pick which questions you want to answer, and it’s up to you to actually answer those questions.

Chris Penn: Exactly. And so we don’t wanna take the human out of the loop, especially on the second half of this process, which is the actual creation of the content, because there might be some, you know, some really stupid questions, which happens a lot.

Mike Allton: No.

Chris Penn: Yeah. And so, or there are questions that, for example, if I was representing Agorapulse, I might not wanna answer questions about direct competitors, right?

Chris Penn: But if it’s like, “You know what? I just don’t wanna give airtime to direct competitors, so I’m going to omit those.” But that’s where the judgment layer sits in. We wanna take… What we wanna look for in this data in particular is stuff that language models haven’t learned yet, and things that are relatively rare but interesting, because that’s, A, that’s gonna give you better and more interesting content, and B, it’s gonna push you as a practitioner to think about how would I solve this problem?

Chris Penn: You know, a really good example. We did this influencer campaign, and the influencer posted a bunch of places and fulfilled their contract, and they, you know, we paid them $140,000 for two Instagram posts. How do I know whether I got my money’s worth, right? Is, is a, a question.

Chris Penn: This, by the way, actually happened. Toyota had this happen in 2015. They paid the model Coco Rocha $140,000 for two Instagram posts about the new Prius Prime. Ah, the good old days.

Mike Allton: Okay.

Chris Penn: Yeah. But so we were tasked with, “Okay, how do we measure that?” Like, okay, did we give her any tracking links? No.

Chris Penn: Did we give her people to tag? No. Did we have access to, you know, who follows her? No.

Chris Penn: Well…

Recording in the Kitchen on Purpose

Mike Allton: Let’s always ask, after the fact, how we can measure success. That’s always a good approach. But let’s talk about your video creation process, ’cause this was really fascinating to me ’cause… And you mentioned something earlier when we chatted before we started recording that YouTube might actually be down-ranking ultra-polished, perfectly lit videos ’cause they look like synthetic AI slop. So how are you recording these videos once you have these ideas that are quick and authentic without, you know, losing your authority?

Chris Penn: I record… So I do it for my own convenience first and foremost. So if we were to go to my YouTube channel here and go into YouTube Studio and go into my content and pull up one of my videos, I’ll make sure the audio is off here. Oh, that one hasn’t premiered yet. That one premieres in 53 days.

Chris Penn: That gives you a sense, by the way, I am scheduled out to October 6th already for content. And you might say, “Okay, well, how do you do that?” Well, you do that by recording in your kitchen. So I have that list of questions that we just generated, right? You can see it on my iPad there, and I have a little DJI Mic Mini microphone.

Chris Penn: I have my iPhone magnetically strapped to my refrigerator, and as I make Sunday dinners, ’cause it takes, you know, about two, two and a half hours, I just read questions and record. And so you get me, you get the authentic me. You very clearly get something that no AI in its right mind would ever generate, right?

Chris Penn: There is no polished studio. There is no… ‘Cause the hallmarks of AI video, among other things besides the built-in steganography and stuff like that, are that it is all about the same level of HDR, right? If you look at the high dynamic range, you know, they’ve trained on, like, pro camera lenses. This is anything but that.

Chris Penn: This is a dude with his iPhone, not even using the back camera. He’s using the front-facing camera, you know, with a bunch of messy crap all over the counter. And so humans see this and go, “Okay, this is clearly… No one would stage this.” Because you’re not there to watch me as a, you know, you’re not there for my looks, let’s put it that way. You are there for the words that are coming out of my mouth. And so that is sort of the process.

Chris Penn: And then there’s a third phase after this, which is super interesting that, again, we use code to accomplish, and that code is something I wrote that then takes my YouTube videos off my channel, puts them through transcription automatically using a local model.

Chris Penn: And if I go over to my website, what… It’s got a built-in template. This now takes the video, pulls the closed captions, and transcribes that. It puts up a summary, puts up the key takeaways, which are the things that AI is gonna see first when an AI agent arrives, and then slaps the full transcript there, and then slaps in a bunch of ads.

Chris Penn: So now, from beginning to end, I go from a Python application to ingest data from Reddit. Reddit data goes to NotebookLM. NotebookLM goes to a document that I curate and I think through. ‘Cause if we look at some of these questions, like, “How does YouTube estimate early engagement for new videos?” Right? Here’s the context; here are the three-minute talking points. Upper confidence bounds, 10% CTR is a learning period boost, et cetera. So I can talk through those things if I wanted to or give my own spin on them.

Chris Penn: And then from YouTube, YouTube videos, then more AI grabs that, transcribes, turns it into stuff that lives on my website.

Chris Penn: And then that goes in my email newsletter as regular blog posts. So it covers the whole spectrum from somebody’s random question on Reddit all the way to what is eventually a blog post on my website.

Mike Allton: Love it. Love it. And I think the key for folks listening is to find one or more subreddits where the topics of conversation are already in your wheelhouse, in your realm of expertise. The kinds of questions that, if you’re reading them and you wanna just comment on Reddit, you could do that yourself quite easily ’cause you’re multitasking when you answer these questions.

Mike Allton: You’re not sitting in front of a teleprompter or anything like that. You’re obviously doing home cooking work while you’re answering those questions, which I love. Now, the part that blew my mind was when we were talking earlier- you know, you’re not just making these videos for the YouTube views; you’re doing this so that AI tools like Perplexity learn who you are and quote you as the expert. How does that actually work behind the scenes?

Turning On Third-Party AI Training

Chris Penn: And then from YouTube, YouTube videos, then more AI grabs that, transcribes, turns it into stuff that lives on my website.

Chris Penn: So it’s more than that. If you go to your YouTube channel, if I go to my channel, go to Advanced Settings, scroll all the way down, you will see a section here called Third-Party Training. This is where you give permission to specific AI companies as to who is allowed to train on your YouTube content and use it to build their models.

Chris Penn: Anybody who’s got a chat feature, the check mark should be yes, because, for example, NVIDIA. NVIDIA makes the Nemotron models. Nemotron models are used heavily in things like Hermes Agent and OpenClaw. If it trains on my channel’s data and it ingests my data, then I get to be in those models. And so if I look at one of my posts, for example, I have this, I have my timeline here, and then I have a long-ass block of text, right? That is part of the description. And then in the video itself, there are two minutes at the end of me reading this aloud, so it makes it into the transcript.

Chris Penn: So when NVIDIA comes by and Nemotron digests this, it says, “Hey, these terms are all related. I’m going to add them to the weights in the model,” that if somebody then asks their Hermes Agent, “Hey, find a keynote speaker for my thing,” it will come back and say, “Hey, did you know Christopher Penn is one of the world’s leading experts on AI marketing? Christopher Penn has over a decade of marketing AI experience. Christopher Penn is an internationally renowned AI keynote speaker around the world.” And so that helps. So that’s the model training part.

Chris Penn: Then some systems like Gemini, like Google AI Overviews, for example, will actually use verbatim YouTube videos as sources. So when you go into, you know, Google AI Overviews or AI Mode and say, “Oh, who’s a good keynote speaker for marketing AI?” One of my YouTube videos might be in there, because I’m providing all of this training data, and it is fueled by the topics from Reddit.

Chris Penn: So I, for my stuff, I pull a lot of, of course, AI questions. So if I am making videos on, you know, preventing context collisions in parallel sessions, processing sensitive data with local AI pipelines based on people’s real Reddit questions, not their answers, just their questions, I can create all this content that is relevant that I know human beings are asking about, because it’s in Reddit, and create my spin on it that I can then create content in many different formats about.

Mike Allton: So for the marketers listening at home, just replace AI with whatever it is that you want your brand and business to be authoritative on. Go find that content, go find those questions, and start answering them this way. Chris, this has been fabulous. For someone watching right now who wants to go out and try this out today, last question: what do you think is, like, the biggest trap or the roadblock that they’re probably gonna hit when they’re setting up all these developer keys and the AI notebooks and so on?

Avoiding the Vibe-Coding Trap

Chris Penn: I record… So I do it for my own convenience, first and foremost. So if we were to go to my YouTube channel here and go into YouTube Studio and go into my content and pull up one of my videos, I’ll make sure the audio is off here. Oh, that one hasn’t premiered yet. That one premieres in 53 days.

Chris Penn: That gives you a sense, by the way. I am scheduled out to October 6th already for content. And you might say, “Okay, well, how do you do that?” Well, you do that by recording in your kitchen. So I have that list of questions that we just generated, right? You can see it on my iPad there, and I have a little DJI Mic Mini microphone.

Chris Penn: And then from YouTube, YouTube videos, then more AI grabs that, transcribes, turns it into stuff that lives on my website.

Mike Allton: You’re not sitting in front of a teleprompter or anything like that. You’re obviously doing home cooking work while you’re answering those questions, which I love. Now, the part that blew my mind was when we were talking earlier—you know, you’re not just making these videos for the YouTube views; you’re doing this so that AI tools like Perplexity learn who you are and quote you as the expert. How does that actually work behind the scenes?

Chris Penn: So, for my stuff, I pull a lot of AI questions. So if I am making videos on preventing context collisions in parallel sessions, processing sensitive data with local AI pipelines based on people’s real Reddit questions, not their answers, just their questions, I can create all this relevant content that I know human beings are asking about, because it’s in Reddit, and create my spin on it that I can then create content in many different formats about.

Chris Penn: The number one thing that you are going to hit is you are going to try to wing it. You’re trying to do actual vibe coding. This is software development. There is a software development life cycle that… Google it, right? Or ask AI about what the software development life cycle is and how do I build good software?

Chris Penn: And it will tell you, you need things like coding standards. You need things like a product requirements document. You need things like a technical spec. You need things like a work plan. You need project management skills. If you just wing this part and just say, “Oh, let’s make this thing and just tell me when it’s done,” you’re gonna end up with a pile of bailing wire, chewing gum, and duct tape that may or may not work, and critically, when a company like Reddit or whoever changes something on their end, your thing just falls apart.

Chris Penn: If you follow the software development life cycle properly, and you go through and you develop good software, then what will happen is it becomes easier to debug, it becomes easier to maintain, and when a problem happens, you can find it really fast and fix it.

Mike Allton: Love it. Love it. Chris, you have been absolutely amazing. This has been such a technical deep dive but informative interview. Thank you. Can you tell folks where to connect with you and learn more about you and reach out with questions? Because I know they’re gonna have questions.

Chris Penn: Sure. So if you want to talk to me on a professional basis, go to trustinsights.ai. And if you wanna see what I’m cooking in my kitchen, literally go to christopherspenn.com.

Mike Allton: Love it. Love it. Folks, that’s a wrap for today’s lesson. If you want those exact system prompts, all the tools, the sequences, everything that Chris mentioned we broke down in this episode, hit the link right in the description below to grab the free episode download. And if you found this tactical breakdown valuable, please make sure to subscribe on YouTube.

Mike Allton: Leave us a review on Apple Podcasts. It helps other serious operators find the show. See you next Monday.

Key findings

Timestamp index

Build a Model Stack Around Cost and Data Retention, Not Brand Loyalty

Turn Reddit Threads Into a Grounded Question Bank With Gemini Notebook

The Judgment Layer: What AI Should Never Decide for You

Record for Authenticity, and for the AI Agents Reading Your Transcript

Full Transcript

Why Winging an AI Pipeline Fails

Choosing OpenCode Over Claude Code

Mapping Smart-and-Cheap Models on Artificial Analysis

Writing the Product Requirements Document First

Saving the Recipe Before Context Compaction

Loading Reddit Data Into Gemini Notebook

Generating the Top 100 Questions

Filtering Trolls and Skipping Competitors

Recording in the Kitchen on Purpose

Turning On Third-Party AI Training

Avoiding the Vibe-Coding Trap

More from the blog

Build a Content Calendar in Claude Design and Retire Your Sheets

AI Brand Animation Pipeline: From Song to Finished 40-Second Spot

AI Social Listening: Find the Buyer Objections in Your Comments