Shrimp Brainstorming: Wie KI-Agenten Projektideen vor dem Coden prüfen

So, today's video is going to be all about brainstorming. I'm going to be talking about how I brainstorm with my agents and the skill that they produced, which is called Shrimp Brainstorming. This is available on my GitHub at TomBiStudio/shrimp-brainstorming. You'll be able to see it for yourself. And this is a skill I developed. Actually, the Hermes agent developed it naturally. I never asked it to do this, but over several months of doing brainstorming sessions with my Telegram Hermes agent, we just kind of came up with this skill that works really well for us. So, in this video, I'm going to kind of try to introduce you to how I brainstorm, my process, and then we're going to go through an actual brainstorming session, which will lead not to a finished product, but to kind of a spec strip, or something that I can hand off to a proper coding agent because many of you have seen in my videos, I start with these kind of handoffs or spec scripts, already a rough plan of what I want to do. So, in this video, I want to take a kind of step back and see, where does that rough plan come from? And it comes from Shrimp Brainstorming. So, let's get started. And if you run agents yourself and want access to the LLM wikis that I use myself for these videos to do research and actually create them, check out my project agentwikis.com. I provide all of these wikis for free on a variety of different topics. And you could also sign up for a Pro account for $9.99 a month that will give you access to supersized, extra-large wikis that will give you more pages and more detail on each of the subjects. Now, back to the video. So as you know, there's a lot of different brainstorming skills out there. Superpowers has a very popular one, agentic awesome skills, there's a ton of them. And I started thinking about brainstorming when I did the grill me from Matt Pocock. And grill me is a little bit different from a brainstorming skill, but it's in the same general idea, I think. So a few things about how I do this because I have kind of a dedicated bot here, and this is Scampi. If you've been watching my channel since February, you've seen me talking to Scampi, which started as an OpenClaw agent, which is why it's a shrimp character. But I migrated it over to Hermes agent, and I've been talking to it through Telegram since then. And Scampi's main role is just to kind of bounce ideas off of. I don't expect Scampi to come up with great ideas. In fact, when I've asked for great ideas, they are usually terrible. So these kind of agents, in my experience at least, are not good for coming up with the ideas themselves. They are good for brainstorming and doing the research I need to kind of flesh out an idea and let me know what's already been done, what already exists, and where potential openings are. And you see for this Telegram agent, who is living on an old laptop, like a 7-year-old laptop that can't do any actual coding on it really except maybe basics, like skill files, and that's why I have it kind of dedicated to this role. And the main model I use for this is GPT-5.6-Terra. You could probably even do Luna, Sol is way overkill for this, and it's too slow. So I find Terra is a good balance. Um, you could also use some of the faster Grok models, I think. I've used them as well, and those are pretty good as well, but I have a lot of Codex resets that I need to use, so I try to get the most out of my GPT-5.6, so I find Terra is a good option there. So this process is pretty simple, right? I come with a real half-formed idea. Sometimes it's a kind of a paper that I've seen, something I've seen on Twitter, probably that looks interesting and I'll say, you know, can we recreate this? What do you think of this? Sometimes it's about some tool I saw, also on Twitter. I'll ask about it. Many times, though, it's just a general idea for some kind of business or project. If you've been following along with my newsletter, AI Garage Weekly, you'll see a lot of my project ideas. They usually don't work out, but I've had ideas for AI-powered custom children's videos, I had an idea to use AI for watches and luxury goods, and I had an idea to use AI to create kind of enhanced quotes for HVAC companies. So those are among the kind of random ideas I've had. And some of them do end up in the videos, like the one you saw last week, or whenever it was, for GitLana. That was basically just me asking my agent if I could put code on the Solana blockchain. Now, 99% of the time, these ideas don't end anywhere, but sometimes they do. It all starts with one question, right? So I'm going to run through this with my agent. I'm going to start a new session here, as we go through this, and I'm saying, I want to brainstorm this idea I had for an AI breadmaker. See, and the first thing it does is skill view shrimp brainstorming. And you may be thinking this is a stupid idea. It is a stupid idea. But my brainstorms are often this stupid, and that's why most people don't see them. But I think the freedom of having an agent, right, is that you can talk about the stupidest ideas you have, which in my case right now is an AI breadmaker. Um, so sometimes it'll ask me clarifying questions, right? Am I talking about an actual bread machine or some kind of like software product? Um, I said it's an AI recipe software product that connects to breadmakers. So now, it'll often do research, and this is kind of the bulk of the brainstorming session, is me asking questions, or in this case answering questions, and it doing research about what it's supposed to do. So this is kind of the flow, right? The user brings a half-formed idea, the agent researches and pressure tests it, user has questions, challenges, findings, and redirects the inquiry, the idea becomes clearer through evidence and decisions, and then a final spec MD records the journey and the recommended next move. So where we are on the research, okay, we got a shrimp maybe. And that's one of the aspects of this skill is you'll get a shrimp verdict, um, here it is, shrimp verdict. So you get either a shrimp yes, a shrimp maybe, or a shrimp no, if the idea is no good. So so far we got a shrimp maybe. But only if we kill the fantasy of a universal connected breadmaker early. So, most breadmakers are deliberately dumb at appliances, there's no connectivity. So, there is a promising wedge here, a breadmaker copilot, kind of like that. So the user can say, make soft Japanese milk bread, 750 grams, less sweet, using what I have. So then the app knows their exact machine, produces a compiled recipe, adapts to local conditions and past outcomes, a diagnosis failure. This is much more defensible than AI makes bread recipes. So, they're saying, don't connect it to the actual appliance, kind of create it as an, a software app that, I express my intent, and then it crafts a recipe based on what it knows about the hardware, the soft, the breadmaker that I have. So this is a good idea. Paste a URL screenshot cookbook recipe, Tik-Tok. The app extracts it, converts units and loaf size, uh, flags ingredients. So this is actually a decent idea. Probably can't connect to the breadmaker, but let me ask that. Can any breadmakers connect via Wi-Fi to apps? It says that most of them can't, but I wonder if that's true. It says that, effectively, no. At least not in a mainstream, interoperable way. So it seems like not connect to a breadmaker directly. So it's saying, do not build software that connects to breadmakers, build AI software that understands breadmakers. I guess it's not the idea I had. Okay, so let me ask another question here. And you see a patched SkillMD, and this is Shrimp Brainstorming. So this is how this skill is kind of developed over the months, and it's not a terribly complicated skill. Um, you can see, check the SkillMD file here. This is what it does. So this is when to use, right? And as a vague products, service, startup, creator, venture, tool, or business idea. So, conversational model. This is user-led inquiry, not a mandatory question funnel. So that's the idea, right? Um, we did borrow a couple of these ideas from other brainstorming skills that I've done research on. Something like Fog of War, in-scope uncertainty, that is not yet clear enough to investigate precisely. Um, here's the general workflow as well. You're interested, orient, research the reality, find the wedge, which we tried to do right here. You've seen this with the breadmaker example, give the Shrimp verdict, and then maintain a living record. And that's the kind of handoff, which is not only the final idea, but it's the whole journey of how you get here. And how the idea evolved. Uh, so the next question, is there any software like this? And this is a very common question I ask. I ask about, you know, any competitors, anyone who has already done this. I'm not really interested in just redoing something that someone else has already done. That doesn't really interest me. So I often ask about that. Not exactly. There is a crowded layer of adjacent bread software, but I found no obvious product combining breadmaker model knowledge and recipe conversion and learning from outcomes. So there are a couple different product types, some examples, Bread Boss, that's a pretty great name. Rise, also a great name. Um, so some of these are baking journals, right? Modern baking companion, Bread Formula Tools, uh, recipe catalogues. So the real gap is not AI bread recipes, because there's already kind of stuff like that. Bread engineer seems to be, claims on-device ML to predict fermentation peak. That's pretty cool. Um, so it is the kind of wedge that we're trying to find here is, I own a, you know, X model. Uh, take this random recipe and tell me exactly which buttons to press, which program to select, what must change, and learn from how my loaf turned out. So the learning function is actually interesting because you can kind of journal that, and it can learn from your experience. So this means the product needs a breadmaker profile database, recipe compiler. So, shrimp yes, we got a shrimp yes on the wedge. Not yet, yes on the not yet on the business. It's a real product gap, but the dangerous question is whether enough breadmaker owners care enough to repeatedly use it and pay for it. Uh, which is fair enough of a concern. So I'm asking, how would you build this out technically? A mobile app. I guess that would make the most sense, right? Uh, so says yes, mobile first, but I would not begin with a full native app. Build a kitchen-friendly app, uh, once the core loop is proven. So this is now we're building out the technical shape a little bit here. I also wanted to ask the key technical decision. AI should not decide critical numbers on its own. Uh, so we're going to need a compiler to produce the actual bake plan, which is going to be deterministic. So now it's building out version 0, version 1. Let me ask, uh, how did those other competitor software monetize? The one it found in the research, run. I'm always thinking, you know, if I'm going to make this a business or a product, is there a way I can monetize it in some way? So that's a good, uh, good way to brainstorm as well, because it's connected to this kind of competitor. You see how other software developers kind of monetize and how they build out the products. Okay, so our research came back, they mostly split into two camps, cheap one-time utility or freemium subscription. People are really paying this upfront $6. Um, free with limited recipes, features. And then do kind of a freemium year. Uh, free plan, then Kneady Pro. I have the greatest names for these these bread-related apps. Okay, so it seems like that kind of freemium thing is, is common here. So what it means for do not copy Bread Boss's $5.99 one-time model if the product uses AI image parsing. That model dies the second every recipe import has an inference cost. So the free, one saved breadmaker profile, three recipe conversions per month, manual recipe entry, basic bake plan and a small history. Loaf Pro, uh $5 a month. Later, an optional credit pack. For expensive vision-heavy actions. Um, I think the vision aspect could be useful here. So the idea here is that users subscribe for the reliable personal breadmaking system, not for a recipe generator. So that's also in the kind of this brainstorming skill, is that it tries to find the perspective of a product, um, kind of your angle for it, not just accept the product itself. Okay, so let's see our shrimp verdict here. Uh, we got a shrimp maybe. Okay. As a lean, tasteful niche product, yes. As a big AI startup thesis, no. So it does, it's not like sycophantic, right? It's not just going to be your yes man. It's going to give you the honest uh, shrimp verdict here. Uh, so the correct move, do not build the app yet. Run a concierge test. Paste any recipe and tell us your breadmaker model. We'll send back a machine-ready bake plan, then help you fix the next loaf. Start with five machine models. Do 25 to 50 plans manually or semi-manually. Charge nothing at first, but ask each tester, "Would you pay $30 a year for this if it reliably made online recipes work in your specific machine?" Okay. Okay, so the idea is you get this, you ask enough questions, it does research, you kind of hone in on the idea, right? And you don't need to, like I say, you don't need a full design plan. That is for your coding agent, whatever you're using. The idea is just to kind of scope it a little bit, do research on it, and just think it through. Honestly, most of this process is just you thinking about it. And I know I used kind of a silly example here, but this could be anything. It could be like a technical experiment, it could be something a lot more ambitious than a breadmaker app. It could really be anything. Uh, so lastly you say, write the spec, and this is going to be the handoff file. This is really usually what I end up with. I just have a blank directory, and then whatever spec they give me, whatever brainstorm file, and that's usually how I start my projects. And whether that's, you know, an experiment, kind of a paper recreation, trying to build a product like this, trying to build out a service or some kind of, uh, business idea, uh, this is usually where it all starts. And there you go, you wrote it out. Posted it here in Telegram, and that's part of the skill as well. You can see here there's a template for the final spec, but it, this is the kind of the idea, original spark, destination, the journey, um, open uncertainty. There's usually a lot of uncertainty. And then a validation plan, then recommended MVP or first offer. So you can see our spec for this breadmaker copilot, and like I said, this is the only thing I would hand to a Claude code or, if I'm working in Hermes Agent, uh, whatever coding agent I'm deciding to use. This is what I handed. So it has a summary, the original spark of the idea. The original idea was an AI recipe or software product that would connect to breadmakers. Chronic, chronological brainstorming narrative, the connection premise, the real product wedge, competitor and substitute scans, technical direction, monetization, user problems, so it has all of this part of it, because I feel like an, in a lot of cases, that is just as important. Um, one because the agent can make mistakes, and two, it just kind of shows the coding agent what you're really thinking about. What is most important in this project, um, what you really want it to be, um, and it just includes all of that as well as all the research you did for competitors and everything like that. Uh, so then I just start in Claude Code, right? And I say, please read, um, whatever the spec is, and it reads it and then you can start building, right? So that's the process, the brain brainstorming process. And this is what I do all the time, basically. And like I said, I mainly use Scampi here, my Telegram bot. I do it mainly in Telegram because I don't have these ideas when I'm sitting and actually working in my office. I usually have them out and about when I'm doing something completely different. So it's usually at that point then I'm throwing it to my Telegram bot here. Uh, so that's the whole process. And if you'd like to check out my skill, I did open source it. Um, it's at TomBiStudio on GitHub/shrimp-brainstorming. Hopefully, you find it useful. But that's going to be the end of this video. Please leave a comment, let me know how you do brainstorming, if you use any of those kind of skills, or what your process is. I think it's always interesting, uh, because brainstorming is really one of those things that I think is really useful using agents. Uh, but that's going to be for this video. Thank you for watching.