Building Jigs 2.0
The Line Just Moved: What Happens When “Building a Jig” Means Deploying a Real App
This will be the last time you hear from us in 2025; watch for our next issue in early January. Thank you so much for subscribing and sharing this newsletter; we’re ending the year with well over 2,000 subscribers, wildly exceeding our expectations. ❤️
Back in Issue #8, we introduced the concept of jigs: low-investment, custom-built AI tools that solve specific problems. At the time, that meant custom GPTs in ChatGPT, Claude Projects, and Gemini Gems. Chat interfaces with specialized knowledge. Useful, but fundamentally limited to conversations inside the AI platform itself.
It’s wild how much has changed in the months since.
What you can accomplish when “just building a jig” has expanded so dramatically that it barely resembles where we started. Today, you can build a jig that is a fully functional, deployed web application that can be shared with anyone via a link.
After the news round-up, we’ll dig into just how far that line has moved, how you can experience it yourself, and how to enter our inaugural ❄️Winter Jig Competition ❄️.
📰 What’s Happening in GenAI
ChatGPT 5.2
Sam Altman declared a “code red” after the release of Gemini 3.0, and less than two weeks later they released ChatGPT 5.2 alongside rumors of another new model coming in early January. We’re still getting to know this model, but in early testing we enjoy working with this one much more than ChatGPT 5 or 5.1. They’ve clearly worked on its personality and its ability to use tools; it feels much more natural to talk to than its predecessors.
New Standards
Late last year, the team at Anthropic launched “Model Context Protocol,” a structured way for GenAI models to talk to other applications. It exploded in usage, with both Gemini and ChatGPT adopting it even though it was technically owned by Anthropic. In what is a positive sign for interoperability across GenAI models, Anthropic has gifted MCP to the Linux Foundation, which has created a new subsidiary: the Agentic AI Foundation. Avoiding GenAI vendor lock-in is a huge consideration for SMBs; open standards go a long way towards mitigating that concern. Here’s hoping that the next thing Anthropic gives to this foundation are “Skills” 🤞🏼.
The Word of the Year Is…
Slop! Continuing their annual tradition, the good folks at Merriam-Webster have picked this AI term as 2025’s Word of the Year, defining it as, “digital content of low quality that is produced usually in quantity by means of artificial intelligence.” We think this video captures the essence of slop perfectly; enjoy.
“What’s Possible” Keeps Evolving
In January, the frontier of what a non-technical person could build was a custom GPT, Claude Artifact, or Gemini Gem. Helpful, but limited.
By spring, tools like Lovable.dev and Clay.com expanded the boundary. You could create working prototypes and marketing automations, though deployment still required some technical navigation.
By fall, Replit Agent 3 made it possible to go from idea to fully functional, deployed web app just by chatting (and forking over maybe $100). The agent handles everything: writing code, testing, setting up databases, checking security settings, configuring hosting, deploying to a live URL, and even hosting.
Today, you can do all of this basically for free using Google AI Studio’s Build mode.
This signifies a pivotal shift in how users leverage technology to drive outcomes. We’re not quite at the point where you can get any software on demand, but that day is coming soon. We’ve crossed a threshold where certain categories of applications are genuinely accessible to anyone willing to learn a simple process and put in a bit of time.
The “jagged frontier” of GenAI capabilities is going to keep advancing. Exploring its edges today and learning how to build and deploy applications is the best way for you to deeply understand just what’s possible today and get ready for what’s going to be possible tomorrow.
Same Philosophy, Bigger Canvas
The jig philosophy hasn’t changed: build something low-investment knowing it might end up back in the scrap pile. Solve an immediate problem, learn something, and turn yourself into a “savvy buyer” of the right solution. What has changed is the canvas you’re working on.
Today, “just a jig” can mean a deployed application with a database, user authentication, and integrations with other platforms. The disposable mindset stays the same. The capabilities are radically different.
This matters because it should change how you think about what to build. You’re still not creating enterprise software meant to run for years. You’re building functional prototypes that unlock business value right now, and you’re developing new skills so that you can build new jigs whenever you need them.
A Real Example
Last month, I needed a tool to conduct voice interviews about attitudes toward generative AI. I wanted something that could ask questions verbally, capture responses, and store the data for analysis.
Within about an hour, I had a fully working application that used the Gemini Live API for real-time voice conversation, asked interview questions in a natural conversational flow, captured and stored responses, and was deployed to a URL I could share with anyone.
No code written by me. No deployment configuration. Just conversation and iteration with GenAI until the app worked.
Will I still be using this exact tool in six months? Nope. Heck, I probably won’t ever use this same app again. The capabilities will have evolved, my needs will have shifted, and I’ll build something better. But that’s fine. The tool served its purpose, I learned the process, and I can rebuild something faster and better the next time.
The Three-Step Process
Here’s the approach that works:
Step 1: Design Through Conversation
Start with your preferred frontier model (Claude Opus 4.5, ChatGPT 5.2, or Gemini 3.0) and have a conversation about what you want to build. Use extended thinking or reasoning mode for this phase.
Don’t just describe the end product. Work through it collaboratively:
What problem are you solving?
Who will use this and how?
What’s the user flow from start to finish?
What are the core features vs. nice-to-haves?
Ask the AI to help you create a looks-like/feels-like prototype. In Claude, this means building an Artifact. In ChatGPT, use the Canvas feature. The goal is a non-functional mockup that captures what you’re trying to build.
This conversational design phase is critical. You’re not just planning; you’re pressure-testing your assumptions. The AI will ask clarifying questions that reveal gaps in your thinking.
Prefer something more structured or aren’t quite sure where to start? Check out this very cool PRD-builder that my friend Burton Rast made. If you work better by going through a questionnaire, start here instead.
Step 2: Generate a Detailed PRD
Once you’ve refined your concept through conversation, ask your AI to write a comprehensive Product Requirements Document. Use this exact prompt:
“Write an incredibly detailed PRD (product requirements document) as though you are an employee at Google. Include product summary and goals, user personas and use cases, detailed functional requirements, technical specifications, data models, error handling, and acceptance criteria for each feature.”
The PRD isn’t bureaucratic overhead. It’s the blueprint that will guide the vibe coding tool in the next step. The more detailed and precise your PRD, the better your results. Think of it as explaining your project to a very literal-minded junior developer who will do exactly what you say, nothing more, nothing less.
Review the PRD carefully. Edit anything that doesn’t match your vision. This document becomes the source of truth.
Step 3: Build and Deploy
Now for the magic. You have two paths:
Free Path: Google AI Studio “Build”
Navigate to aistudio.google.com and select the Build tab. Paste your entire PRD into the input box and hit enter.
Watch as Gemini generates your code in real-time. You’ll see files populate in the explorer, a live preview appear, and the AI’s reasoning process unfold. It’s transparent about what it’s doing: analyzing concepts, outlining API integrations, mapping project components, developing app logic.
Once it’s working in Preview (more on how to get there below), click the <deploy> button. Google will deploy your app to Cloud Run, giving you a public URL. The entire process can take under 2 minutes for simple applications.
Key advantages: no cost for experimentation, one-click deployment, and usage by people who view your shared app is attributed to their own free quota, not yours. Note that Gemini will train on this data, so don’t put anything sensitive or personally identifiable. Remember, this is a free tool.
Paid Path: Replit
If you want more sophisticated features (databases, authentication, integrations with Slack or email) as well as more privacy (your data is excluded from training), Replit’s Agent 3 is a better option. It costs $25/month for the Core plan, but you’ll likely need to load up around $100-150 in credits in order to create, test, and deploy a sophisticated app.
Paste your PRD, turn on “Autonomous - Max” and “App Testing” in the settings menu inside of your prompt box, and let Agent work. It will ask if you want to see a prototype or have it just start building. I usually pick the latter. It will test its own code, fix issues it finds, and iterate autonomously for about 20 minutes at a time. I find working with Replit is a few minutes of work a few times an hour; it’s become something I love to do on the side while watching a football game.
When you’re satisfied, hit Publish. It really is that easy.
Replit’s advantages include built-in databases and object storage, the ability to build agents that connect to other platforms, more robust production features, and privacy.
Expect to Iterate
These tools will rarely nail your entire vision on the first pass.
More often, Google AI Studio or Replit Agent will build a substantial chunk of your app, then pause and ask for feedback. Or it will complete something that mostly works but misses details. This is normal. This is the process.
Your job at this point is to click through the app in the preview window. Test every button. Try every flow. Identify what’s working and what isn’t. Then give the model specific feedback: “The submit button doesn’t save to the database,” or “The confirmation message appears before the form validates,” or “Add a loading spinner while the API call runs.”
The detailed PRD ensures the AI is building the right thing in the right direction. It has the holistic picture. But from our experience, even a comprehensive PRD won’t capture every nuanced detail of how you want the app to behave. That’s what the iteration loop is for.
This human-in-the-loop feedback cycle isn’t a failure of the tools. It’s how vibe coding actually works. You’re collaborating with AI to refine something into existence, not ordering a finished product from a menu.
What Makes This Work
Three things separate success from frustration:
Clarity of intent matters more than technical knowledge. The AI can handle the technical implementation. What it can’t do is read your mind about what you actually want. Invest time in the design conversation and PRD.
Start simple. The best vibe coding advice I’ve encountered: “Solve small problems. Don’t try to build it all with one prompt.” Even with a detailed PRD, complex applications benefit from iterative building. Get the core working first, then add features.
Give specific feedback. When something doesn’t work, describe the problem precisely. “It’s broken” helps no one. “The date picker allows past dates but should only allow future dates” gives the AI exactly what it needs to fix the issue.
❄️ Winter Jig Competition ❄️
Pick something to build that meets three criteria:
Creates immediate value. Don’t build something speculative. Build something you’ll actually use in the next month. A client intake form. A simple dashboard. A tool that automates one tedious thing.
Keeps a human meaningfully in the loop. This isn’t about building autonomous systems. Build something where you’re still making decisions, but the tool handles the tedious parts.
Doesn’t need to last forever. This is key. You’re not building enterprise software. You’re building a jig—something that solves a problem right now and might be obsolete or rebuilt in three months. That’s fine. That’s the point.
The real value of this exercise isn’t the app you build. It’s learning the process. Once you’ve successfully gone from idea to deployed application once, you can do it again. And again. The capability compounds in ways that a single tool never could.
Go through this process. Design it through conversation. Generate the PRD. Build it in Google AI Studio or Replit (or whatever tool you’re most comfortable using!). Iterate until it works. Deploy it.
🏆 How to Enter 🏆
Send us an email at winterjigs@remixpartners.ai by midnight on December 31st with “Winter Jig” in the subject line. Include a link to a 2-3 minute video of your app in action (Loom works great for this), tell us how you built it, and a bit about what you learned along the way.
We’ll pick the most interesting jig, with a panel of ChatGPT 5.2 + Gemini 3.0 + Claude Opus 4.5 serving as our judges. We’ll give them each the full text of this post + your submissions and ask them to rank everything. The winner gets a limited edition Remix Partners hat and will be announced and featured in a January edition of AI for SMBs.
We want to see what’s possible when people who understand their business problems can directly build solutions. Show us what you’ve got 🙂.
The Bigger Picture
One of the things we most love about generative AI is that it democratizes who can control technology to accomplish their goals. It reduces the amount of technical complexity you need to understand in order to command machines.
We’re still far from a world where anyone can get any software on demand. But 2025 has been the year where certain categories of software became genuinely accessible to non-technical builders.
And this line keeps moving. Early this year, that line was “helpful chat assistant with custom knowledge.” Today, it’s “deployed web application with real functionality.”
By this time next year? I genuinely don’t know. But if you want to stay ahead of where that line is going, the best thing you can do is push against it right now.
Happy building. Happy holidays.
Questions? Email us at info@remixpartners.ai - we read every message.