Imagine having a tireless, creative partner that can generate a hundred business ideas before your morning coffee even gets cold. I’m not talking about those generic, surface-level suggestions either. I mean deep, structured, and occasionally wild concepts that actually hold water. Today’s prompt from Daniel is about using AI as a structured brainstorming partner for business ideas, side hustles, and career pivots. We’re going to get into the weeds on technical configuration, model selection, and the frameworks that move this from a toy to a legitimate professional tool.
Herman Poppleberry here. This is a great one because the landscape has shifted so much just in the last few months of early twenty-six. We’ve moved past the era of just asking a chatbot, "give me a business idea," and getting back something like, "start a dog walking service." We are now in the era of specialized reasoning models and multi-agent frameworks that can actually stress-test an idea before you even spend a dime.
It’s funny you mention the dog walking thing because that’s exactly what people expect when they first try this. They get bored because the AI plays it safe. But as we’ll discuss, that’s usually a configuration error or a lack of process, not a limitation of the intelligence itself. By the way, today’s episode is powered by Google Gemini three Flash. It’s writing our script today, so if we sound particularly enlightened, you know who to thank.
Or who to blame if we get a bit too philosophical. But seriously, the core of what Daniel is asking is how to move from generative chat to agentic doing. It’s about building a system that challenges your assumptions. If you’re using AI to brainstorm and it’s just agreeing with you, you’re doing it wrong. You want a system that explores what we call the adjacent possible—those opportunities sitting right at the edge of your current skills or market trends that you can’t quite see yet because of your own cognitive biases.
Right, so we aren’t just looking for an idea generator. We’re looking for a sparring partner. Someone, or something, that can look at a bioinformatics engineer and say, "hey, your specific knowledge of genomic data sequences actually makes you the perfect candidate to disrupt the boutique pet supplement market." That kind of non-obvious connection is where the real money is.
Well, not exactly—Corn, you know I’m trying to stop saying that—but you’ve hit on the divergent thinking aspect. Let’s start with the technical foundation because if you don't get the model selection right, the rest of the framework falls apart. For the divergent phase, where you want a high volume of "crazy" or out-of-the-box ideas, you need models with high reasoning capabilities but also enough latent space to be creative. In the current market, Claude three point seven Sonnet, which dropped back in February, is really the gold standard for this.
Why Sonnet over something like the ultra-heavyweight models? Is it just a speed and cost thing for batching?
It’s partly the architectural balance. Sonnet three point seven has this specific tuning for nuanced reasoning without being as pedantic as some of the larger models. When you’re brainstorming, you want the model to take "leaps." If the model is too grounded in its safety training or "common sense" logic, it will filter out the weird ideas before it even shows them to you. You actually want a bit of that "hallucination" energy, but directed. Think of it like a jazz musician—you want someone who knows the scales but is willing to play the "wrong" note if it leads to a new melody.
So you're saying we should lean into the weirdness. What about the configuration? I know you’re a stickler for the API settings. If I’m sitting in a playground or building a custom harness, what am I sliding those toggles to?
This is crucial. For ideation, you have to crank the temperature. Most people leave it at the default zero point seven. For brainstorming, you want to push that to zero point eight or even one point zero. Temperature, effectively, controls the randomness. At one point zero, the model is much more likely to choose a less probable next token, which in human terms means a more surprising idea. I also like to set top p to zero point nine five. This ensures you still get coherent English, but you're sampling from a wider pool of potential concepts.
And what about frequency penalty? I’ve found that if I ask for fifty ideas, by idea number twenty, the AI starts repeating itself. It gets into a loop of "AI-powered X" or "Sustainable Y."
Frequency penalty is your best friend there. Setting that to around zero point five tells the model, "hey, if you’ve already used the word 'platform' or 'subscription' ten times in this response, try something else." It forces the model to dig deeper into its training data to find different linguistic structures, which often leads to different business models. Instead of a "platform," it might suggest a "guild," a "micro-consortium," or a "drop-ship collective." Those words carry different structural implications for a business.
But wait, how does that work in practice? If I change the word from "platform" to "guild," does the AI actually change the underlying mechanics of the business idea, or is it just a thesaurus swap?
That’s the magic of large language models. The tokens are linked to conceptual clusters. If the model is forced to use the word "guild," it starts pulling in related concepts like "shared ownership," "apprenticeship," and "collective bargaining." Suddenly, your "Uber for plumbers" idea transforms into a "Plumbing Guild with fractional tool ownership." It’s a completely different economic model because the linguistic constraint forced a conceptual shift.
That’s a pro tip. Now, let’s talk about the prompt itself. We’ve moved beyond the "system prompt" being just "you are a helpful assistant." If I want to find a side hustle that doesn’t make me want to pull my hair out, how am I framing that?
You have to use what we call "Few-Shot Ideation" and "Chain-of-Thought" prompting. Don’t just ask for ideas. Give the model three examples of "weird" but successful businesses. Mention things like the Million Dollar Homepage or those companies that rent out goats to clear brush. This sets the "level of crazy" you’re looking for. Then, use Chain-of-Thought. Ask the model to first analyze the intersection of two unrelated fields—say, legal tech and urban gardening—and explain the friction points in that intersection before it suggests a business.
I love that. It’s like making the AI show its work. If it can’t explain why a "legal tech app for community gardens" is necessary, the idea it generates will be fluff. But if it identifies that "land use permits for non-permanent structures are a nightmare for local garden leads," then the business idea it generates—a permit automation tool—has a foundation in reality.
It’s the difference between a dream and a plan. And for those looking at career pivots, there’s a specific framework I’ve been researching called the Ikigai Pivot. You feed the model your current resume, but you also feed it a "passion dump"—just a raw text file of things you actually enjoy doing on weekends. Then you add your salary requirements. You tell the model to act as a career strategist and identify "bridge roles." These are roles that use eighty percent of your current skills but move you into a completely different industry that aligns with your passions.
That’s actually really practical. I think a lot of people feel stuck because they think a pivot means starting from zero. But the AI can see the cross-functional nature of skills that we might miss. A project manager in construction has a lot of the same logistical headaches as a producer in a high-end VR studio. The AI sees the "logistics" and "stakeholder management" as the common denominator.
It’s pattern matching on a global scale. But let’s step it up a notch. If you really want to do this well, you shouldn't be using a single prompt. You should be using a batch ideation harness. This is where you set up a multi-agent system. Think of it like a virtual boardroom. You have Agent A, the "Divergent Thinker," whose only job is to come up with one hundred raw, unfiltered, borderline-insane ideas.
And I’m guessing Agent B is the one who tells Agent A that most of those ideas are terrible?
Precisely. Well—there I go again. Yes. Agent B is the "Cynical Critic." You prompt it to be a skeptical venture capitalist or a tired, overworked consumer. Its job is to find three reasons why every single one of those hundred ideas will fail. It looks for market saturation, technical impossibility, or just lack of demand.
That’s the "Market Reality" prompt you were telling me about. It forces the system to kill its darlings. I think that’s where most human brainstorming fails. We fall in love with our first decent idea and stop looking. The AI doesn’t have an ego. It’ll kill ninety-nine ideas in a second and not feel bad about it.
And then you have Agent C, the "Architect." It takes the three ideas that survived the critic and builds a one-week execution roadmap. One of the best frameworks for this is the "One-Week Minimum." You tell the AI: "Design a version of this business that I can launch with zero dollars and ten hours of work in exactly seven days." It forces the AI to stop suggesting "build an app" and start suggesting "set up a landing page and run twenty dollars of targeted ads to a specific subreddit."
But how do you handle the technical handoff between these agents? If I’m not a coder, can I still run this "boardroom" or do I need to be comfortable with Python?
You can actually do this manually in a single long-form chat window, though it’s tedious. You just say "Now, switch personas. You are no longer the Ideator; you are now the Cynical Critic. Review the list above." However, in early twenty-six, the no-code agent builders have become so good that you can just drag-and-drop these roles into a workflow. But even if you’re just copy-pasted between tabs, the psychological separation of the roles is what matters. It prevents the "Ideator" from self-censoring.
That’s the shift from "thinking" to "doing" that Daniel mentioned. It’s about creating a bias toward action. I’ve seen people use tools like LangChain or AutoGen to automate this whole flow. You press a button, and twenty minutes later, you have a PDF on your desk with five validated, critiqued, and planned business models. It’s like having a miniature McKinsey in your laptop.
It really is. And the cost is negligible. If you’re using something like GPT-four-o-mini for the filtering and critiquing phase—which was released late last year—you can process thousands of ideas for pennies. You use the expensive, high-reasoning models like Gemini or Claude for the initial "thinking" and the final "architecting," but use the "small and fast" models for the high-volume sorting.
Let’s talk about that career pivot angle a bit more. I think a lot of our listeners are in that twenty-twenty-six remote job market where things are getting competitive. How can someone use this to find an edge?
One technique is "Reverse Background Checking." We actually touched on this in an earlier episode, but in the context of ideation, it’s about using the AI to scan the job market for "emerging friction." You can feed the AI recent earnings calls or industry news from a sector you’re interested in—say, renewable energy storage. You ask the AI to identify the "boring" problems. Don’t look for the "build a better battery" problem; look for the "how do we manage the insurance liability of residential lithium-ion installs" problem.
Ah, the "Selling Shovels" approach. Everyone wants to find the gold, but the real money is often in the logistics and the headaches surrounding the gold.
Always. And the AI is incredibly good at spotting those second-order effects. If X technology becomes popular, then Y problem will inevitably arise. If everyone starts using autonomous AI agents for their personal scheduling, then we’re going to need a "Proof of Human" verification service for high-stakes meetings. That’s a business idea that didn't exist three years ago but is becoming a massive friction point now in twenty-six.
I love the idea of using the AI to find the "boring" problems. We all want to be the next big thing, but a side hustle that solves a specific, annoying administrative hurdle for small business owners is way more likely to succeed than a "social media for cats" app.
Unless it’s "AI-powered liability insurance for social media cats." Then you might have something. But seriously, the "Thinking Mode" in models like Gemini is a game-changer here. When you turn on that reasoning trace, you can actually see the AI debating with itself. It might say, "I initially thought about a subscription box for gardeners, but the logistics of shipping soil are a nightmare, so instead, I’ll suggest a digital diagnostic tool for soil pH levels." Watching that "inner monologue" helps you as a human learn how to think about business logic.
It’s like a masterclass in entrepreneurship just by reading the logs. Now, for the people who are ready to try this tonight, what’s the first step? Do they just open a chat window, or should they be looking at these specialized platforms like Taskade or Juma?
Start simple. Don’t get overwhelmed by the tooling. Open Claude three point seven Sonnet or Gemini. Set your persona. Tell it: "You are a world-class entrepreneur with a background in both software and blue-collar service businesses. I want to generate twenty side hustle ideas that leverage my specific skill in—blank." Then, and this is the key, give it constraints. "None of these can require more than five hundred dollars to start. None of these can require me to be on the phone during nine-to-five hours. All of these must have a path to automation within six months."
Constraints are the secret sauce. Without them, the AI just gives you the average of all business ideas on the internet. With them, it has to get creative. It’s like that old saying—the more you limit the canvas, the more the artist has to innovate.
And if you want to go deeper, look into "Recursive Prompting." This is where you take the output of one session and feed it back in with a twist. Take those twenty ideas and say, "Now, rank these by 'Ease of First Sale.' Take the top three and combine them into one hybrid business model." That’s how you get things like "A subscription-based drone service for cleaning hard-to-reach industrial gutters that also generates solar-ready roof maps."
That sounds like a legitimate business. I might have to steal that one. But it really highlights how the AI can synthesize ideas in a way that our brains aren't naturally wired for. We tend to think in silos. The AI just sees data points and connections.
It’s also worth mentioning the "Ikigai" framework for career pivots again. In twenty-twenty-six, we’re seeing a lot of people using AI to identify "AI-Adjacent" roles. For example, if you’re a teacher, you might not want to be a coder, but you’re perfectly positioned to be an "AI Curriculum Designer" or a "Human-in-the-Loop Specialist" for educational tech companies. The AI can help you rewrite your narrative so you aren't "leaving teaching," you’re "evolving into AI-enabled pedagogy."
It’s all about the reframe. And honestly, that’s probably the most valuable thing the AI does. It provides a different lens. We’re so close to our own lives and careers that we can’t see the forest for the trees. The AI is standing at thirty thousand feet with a pair of high-powered binoculars.
And a map of every other forest in the world. But we have to be careful about "Idea Overload." This is the dark side of AI ideation. It’s so easy to generate a thousand ideas that you end up with "Creative Paralysis." You spend all your time prompting and none of your time building.
I’ve been there. It’s the digital version of buying a bunch of notebooks and never writing in them. How do we avoid that? How do we move from the "batch ideation" phase to the "actually making money" phase?
That’s where the "Critic" agent and the "Architect" agent are non-negotiable. You have to commit to a process where the system narrows itself down. If you generate a hundred ideas, you must force the system to give you only one to act on. And then you follow that "One-Week Minimum" roadmap. If you can’t get a signal of interest in seven days, you kill the idea and move to the next one on the list. The goal is to fail fast and fail cheap.
Can you give me a concrete example of that "One-Week Minimum" in action? Let’s say the AI suggests a niche newsletter for people who collect vintage mechanical keyboards. What does the AI tell me to do on Tuesday?
Tuesday is "Landing Page and Lead Gen." The AI would tell you: "Do not write the newsletter. Instead, use a tool like Carrd to build a one-page site that says: 'The Weekly Click: Every Tuesday, the rarest mechanical keyboard finds delivered to your inbox.' Then, go to the mechanical keyboard subreddit and find three unanswered questions about vintage models. Answer them thoroughly, and in your signature, link to your landing page. If you don't get 50 signups by Thursday, the market isn't biting. Move on."
That’s brutal but necessary. It’s the "Lean Startup" on steroids. Instead of taking six months to realize an idea is bad, you take six hours. That compressed feedback loop is the real competitive advantage in the current market.
And for those looking for "crazy" ideas specifically—the kind that Daniel mentioned—don't be afraid to use multimodal models like Qwen two point five VL. These models can "see" and "draw." You can ask them to brainstorm a new physical product and actually have them sketch out the basic UI or the form factor. Seeing a visual representation of a "crazy" idea often makes it feel more real and helps you spot physical design flaws that a text-only model would miss.
That’s a great point. A "smart lunchbox that automatically reorders your favorite snacks" sounds okay in text, but when the AI draws it and you realize it needs a battery, a cooling system, and a cellular connection, you start to see the manufacturing hurdles. It brings the "crazy" back down to earth.
Or it helps you lean into the engineering challenge. The point is, in twenty-twenty-six, the tools are there to be more than just a sounding board. They are a collaborative workspace. Whether you’re using Taskade to build a multi-agent "think tank" or just having a deep, reasoning-heavy conversation with Gemini, the barrier to entry for "The Next Big Thing" has never been lower. But the bar for "Good Execution" has never been higher because everyone has access to these ideas now.
So the differentiator isn't the idea itself—it’s the curation and the speed of validation. The person who uses AI to find the "adjacent possible" and then uses an agentic workflow to launch a MVP in a weekend is going to beat the person who spends three months writing a business plan the old-fashioned way.
Every single time. We’re moving toward a world where "Founder" is more of a "System Orchestrator" role. You’re orchestrating agents to research, agents to build, and agents to market. The "ideation" part is just the first domino.
I have to ask, though—does this take the soul out of entrepreneurship? If the AI is finding the problem, critiquing the solution, and building the roadmap, what’s left for the human?
The taste. The AI can give you a thousand options, but only you can decide which one feels right for your life. It can identify a gap in the lithium-ion insurance market, but if you hate insurance and find it soul-crushing, you won't have the stamina to see it through. The human’s job has shifted from "generating the spark" to "choosing which fire to tend." You provide the intuition and the ethical guardrails.
Well, I’m feeling inspired. I might go prompt a side hustle for a sloth who specializes in deep-dive technical analysis. Oh wait, I already have this job.
And you’re doing a great job at it, Corn. But seriously, for the listeners, this week’s challenge is simple. Pick one interest—something you actually like, not just something you think will make money—and use a high-reasoning model with a temperature of zero point eight to generate ten "weird" business ideas around it. Then, ask a second agent to find three reasons why each will fail. See what’s left standing.
And if nothing is left standing, that’s also a win. You just saved yourself a lot of wasted time. This has been a fascinating deep dive. It’s amazing how much the "how" of prompting changes the "what" of the output.
It’s all in the configuration. If you treat it like a search engine, you’ll get search results. If you treat it like a partner, you’ll get a partnership.
Well said, Herman Poppleberry. I think we’ve given people plenty to chew on. From model selection to the "Ikigai Pivot," there’s a whole world of structured creativity out there if you know which toggles to flip.
And if you're interested in the technical side of building these agentic systems, we’ve got some related episodes in the archive. Episode one-forty-seven, "The Art of the Prompt," is a classic that covers the foundational engineering for creativity—it’s a great companion to what we talked about today regarding ideation frameworks.
Definitely worth a listen if you want to master the "System Prompt" side of things. I think we should wrap it there before you start explaining the math behind top-p sampling again.
Hey, people like the math! It’s all about the probability distribution of the next token. If you don't understand the nucleus sampling, you're just throwing darts in the dark.
We’ll save the nucleus for the after-show. Thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a big thanks to Modal for providing the GPU credits that power the generation of this show. We literally couldn’t do this without that serverless magic.
This has been My Weird Prompts. If you found this episode useful, the best thing you can do is leave us a review on Apple Podcasts or Spotify. It’s the main way new people find the show, and we really appreciate the feedback.
You can find all our past episodes and the RSS feed at myweirdprompts dot com. We’ll be back next time with whatever weirdness Daniel sends our way.
See you then.
Stay weird.
Wait, Corn, before we go—I just realized we didn't touch on the "Inverse Prompting" technique for finding side hustles. It's too good to leave out.
You mean where you ask the AI what it would build if it were a human with only five hours a week?
But specifically, you ask it: "What is a task that humans are currently doing that is so incredibly boring that an AI could do 90% of it, but still requires a human to sign off on the final 10%?" That 10% is your business. You become the 'Human-in-the-Loop' for an automated process.
Like a specialized AI-auditor for medical billing or something similarly dry?
Precisely. It’s the ultimate low-effort, high-margin side hustle because the AI does the heavy lifting, and you just provide the legal or professional accountability. It’s basically arbitrage on your own expertise.
Okay, now we can wrap. That’s a million-dollar tip right there.
Just doing my part for the economy.
Goodbye, everyone!
Bye!
