[Prompt]
Custom topic: A showcase and exploration of the surprising non-development uses people are finding for AI terminal agents like Claude Code, Gemini CLI, and Codex. While these tools are marketed squarely at software developers, a growing community is using them for everything BUT coding — and the results are fascinating.

Seed the discussion with these real examples from the wild, drawn from an open source collection at github.com/danielrosehill/Non-Code-Claude-Code:

FINANCE: Equity research workspaces where Claude Code analyzes stocks and generates reports. LEARNING: Study assistant patterns where the terminal agent helps students research and review material. MULTI-AGENT IDEATION: Think tanks of AI agents for research and policy ideation, panel debates between multiple AI personas, and "Change My View" style deliberation frameworks — all orchestrated from the terminal. PERSONAL PRODUCTIVITY: Household budget management, diary and time planning, even therapy session tracking — all using repo-as-workspace patterns with CLAUDE.md files providing context. RESEARCH: Deep research workflows where Claude Code iteratively investigates topics like an ArXiv research agent, ADHD drug access research notebooks, and technology stack evaluation workspaces. TOOLS & UTILITIES: File system organization, PR and media monitoring, website update notifications to colleagues, structured documentation of system fixes. WRITING & CONTENT: Blog management as a conversational CMS, writing squad agent crews, and general writing workspace templates.

The pattern that emerges: a git repository isn't just for code — it's a workspace container. A CLAUDE.md file becomes a briefing document. MCP servers become tool integrations. The repo structure becomes scaffolding for ANY kind of AI-assisted work.

Discuss why these use cases are flourishing despite zero marketing support from the tool makers. What does this tell us about the actual demand for terminal-based AI assistance beyond coding? Are we seeing the early days of a much bigger shift in how people interact with computers — where the terminal becomes the universal AI workspace? What's holding back broader adoption by non-developers, and could these tools be made more accessible to sysadmins, researchers, writers, analysts, and knowledge workers who would benefit enormously but don't think of themselves as the target audience?

[Response]
Corn: Claude Code was built for developers, but the most interesting uses are coming from people who have never written a line of code in their lives. We are seeing this massive, quiet explosion of terminal-based AI agents being hijacked by researchers, finance analysts, and writers. It is like everyone was told this was a specialized wrench for mechanics, and then a week later, half the neighborhood is using it to cook dinner and redecorate their houses.

Herman: It is a classic case of the users defining the product better than the marketing department ever could. I am Herman Poppleberry, and today we are diving into why the command line is suddenly the hottest productivity workspace for people who do not even know what a compiler is. Today's prompt from Daniel is about the "Non-Code" movement surrounding tools like Claude Code and Gemini CLI. He pointed us toward a repository on GitHub called Non-Code-Claude-Code that has amassed over two hundred real-world examples of this in just about six months. By the way, today's episode is powered by Google Gemini Three Flash, which is actually quite fitting considering we are talking about these exact types of models stepping into the driver's seat of our operating systems.

Corn: It is funny you say "driver's seat" because for most people, the terminal feels like the trunk of the car. It is dark, it is cramped, and you only go back there when something is broken. But Daniel’s prompt highlights this shift where the terminal is becoming the universal AI workspace. These tools, like Claude Code, which Anthropic released late in twenty-four, or the Gemini CLI that followed in early twenty-five, they were sold as "coding agents." They are supposed to help you refactor Javascript or find bugs in your Python scripts. But instead, we are seeing people use them for equity research, ADHD drug access studies, and even personal therapy tracking.

Herman: The technical reason this is happening, and why it is so much more powerful than just chatting with an AI in a browser, comes down to three specific pillars. First, there is the "Repo-as-Workspace" pattern. In a traditional AI chat, you upload one or two files and hope the context window does not forget them. In a terminal agent workflow, the AI lives inside a folder. That folder is a Git repository. To the agent, that repository is its entire world. It can see every PDF, every markdown note, every spreadsheet, and every sub-folder. It has "File System Agency," meaning it can actually move things around, create new documents, and organize your data without you having to click and drag a single icon.

Corn: Right, and that leads to the second pillar, which is the CLAUDE dot M-D file. For the developers listening, you know this as the "instructions" file that tells the agent how to handle the code. But for these non-coders, it is a "Briefing Document." If I am an analyst, I put a file in that folder that says, "You are a senior equity researcher. Your goal is to find debt-to-equity ratios in these fifty PDFs and tell me who is over-leveraged." Every time I open that terminal, the agent reads that file and immediately knows its job. There is no "re-prompting" every session. It has persistent project memory.

Herman: And the third pillar, which is really the secret sauce, is the Model Context Protocol, or MCP servers. This is what lets a terminal-based agent reach out and touch the real world. You can connect it to Google Drive, your Slack channels, or even Brave Search. So now, instead of just a chatbot, you have a command center that can pull data from the web, cross-reference it with your local files, and then write a report directly into your folder.

Corn: I love the idea of a "Panel of Claudes." One of the examples Daniel shared was using the terminal to orchestrate think tanks. You basically spin up multiple instances of the agent, give them different personas in their respective instructions files, and have them debate policy proposals or business strategies. It is like having a board of directors that costs you about twelve cents an hour in API credits and never asks for a coffee break.

Herman: That multi-agent ideation is actually a very sophisticated use of the terminal's architecture. When you are in a GUI, you are usually limited to one "thread" or one "conversation." In a terminal, I can have five tabs open, each running a different agent with a different perspective, all looking at the same set of files in a shared repository. One agent might be the "Devil's Advocate" whose job is to find flaws in a research paper, while another is the "Synthesizer" trying to merge two conflicting ideas. Because they are all operating on the same file system, they can literally "hand off" work to each other by writing to the same markdown files.

Corn: You mentioned the research aspect, and I think the ADHD drug access case study is a perfect example of why this beats a web browser. If you are a researcher looking at fifty different academic papers, trying to find a specific pattern across all of them, doing that in a web UI is a nightmare of "uploading, waiting, asking, repeating." In Claude Code, you just point it at the folder and say, "Iteratively investigate these papers and build me a knowledge base." It goes through them one by one, summarizes the key findings into individual files, and then creates a master index. It is acting like a junior research assistant who actually follows directions.

Herman: The "iterative" part is the key. Most people think of AI as "Input-Output." You ask a question, you get an answer. Terminal agents operate on a "Loop." They can run a command, see the result, realize they need more information, search the local files again, and keep going until the task is done. That is why we are seeing it flourish in finance. An equity research workspace can be set up where the agent "crawls" a folder of ten-K filings. It is not just reading; it is extracting specific data points like debt ratios or revenue growth across different fiscal years and formatting them into a comparative report.

Corn: I have to ask though, Herman, why are people doing this in the terminal? If I am a finance guy or a student, the terminal is usually the place where I accidentally delete my operating system because I followed a bad tutorial on Stack Overflow. Why are these "Non-Coders" flocking to a command line interface when there are perfectly good apps with buttons and shiny icons?

Herman: It is about the "Invisible Demand" for agency. Web UIs are "sandboxed." When you use a web-based AI, it is like talking to someone through a glass partition. They can see what you show them, but they cannot reach through and touch your stuff. A terminal agent has its hands on the actual machinery of your computer. If you want to rename a thousand images based on the content of the photo, a terminal agent can do that in seconds. A web AI would ask you to upload them one by one.

Corn: Plus, there is the privacy angle. If I am using the terminal to keep a personal diary or track therapy sessions—which was one of the weirder, more intimate use cases in that repo—I am keeping that data local. It is in a folder on my hard drive. Sure, the text goes to the model provider for processing, but the "source of truth" is a local Git repo. It is much more private than having a "Chat History" synced to a cloud account that anyone with my password can read.

Herman: There is also a "Watch and Learn" effect that users are reporting. This is fascinating to me. When you tell a terminal agent to "Organize my downloads folder by file type and date," you actually see the commands it uses. You see it run `mkdir`, `mv`, `find`. For a non-coder, this is like having a tutor who shows their work. You start to realize, "Oh, I do not need a special app to organize files; I just need to know these three commands." It demystifies the computer.

Corn: It is like the AI is a translator for the "Machine Language" we have all been too scared to learn. But let's look at the friction here. If this is so great, why aren't my parents using Claude Code to manage their grocery lists yet? Because right now, setting this stuff up feels like building a nuclear reactor just to toast a piece of bread.

Herman: You are not wrong. The "Dependency Hell" is real. To run Claude Code or Gemini CLI, you often need to install Node dot J-S, configure API keys, set up environment variables, and understand basic terminal navigation. For a professional researcher, that is a hurdle they might jump over. For a casual user, it is a brick wall. We are currently in the "MS-DOS" era of AI agents. The power is there, but the interface is still demanding that you meet it halfway.

Corn: And the marketing is not helping. Anthropic and Google are obsessed with "Coding Benchmarks." They want to show that their model can solve "LeetCode" problems or fix a broken Javascript library. They are not measuring how good the model is at "Summarizing a legal deposition" or "Managing a household budget." There are no "Research Benchmarks" for these agents yet. So if you are a writer or an analyst, you look at the landing page for these tools and think, "This is not for me."

Herman: Which is a shame, because the "Writing Squad" use case is brilliant. Imagine a repository where your "CLAUDE dot M-D" defines a squad of agents: an editor, a fact-checker, and a creative director. You write your draft in a markdown file, and you tell the agent, "Run the squad on this." It passes the file through those different personas, each leaving "comments" or making tracked changes in the file system. It transforms writing from a lonely task into a collaborative process with a local team.

Corn: It is basically a conversational C-M-S. I saw that in the repo too—people managing their blogs this way. Instead of logging into WordPress and fighting with a block editor, they just talk to the terminal. "Hey, update the 'About Me' page and make sure the new post is linked in the sidebar." The agent handles the file updates and the Git push. It is "Vibe Coding" but for "Vibe Content."

Herman: I think we need to talk about the "Repo-as-Workspace" concept more deeply, because that is the fundamental shift. Historically, we thought of a "folder" as a place to store things. But in this new world, a folder is a "Context Container." If I am a student studying for the Bar Exam, I create a repo called "Bar Exam Prep." I put all my notes, practice tests, and textbooks in there. Because the AI is "contained" in that repo, it has a boundary. It knows that when I am in this terminal window, I am a law student. It does not get confused by my grocery list or my hobby of collecting rare sloth photos.

Corn: Hey, that is a very respectable hobby. But I see your point. It is "Spatial Organization" for the mind. In a web browser, every tab looks the same. In the terminal, the "path" tells you exactly where you are and what the AI's "brain" is currently tuned to. It is a way of managing the AI's "Attention" by physically limiting what it can see.

Herman: Wait, I promised I would not say that word. It is a way of managing the AI's attention. What is even more interesting is how this enables "Deep Research Pipelines." There was a case study about a researcher creating a "Technology Stack Evaluation" workspace. They would give the agent a broad goal, like "Compare three different database architectures for a high-frequency trading platform." The agent would then iteratively search for papers, download them using an MCP search tool, read them, and build a comparison matrix. It is not just answering a question; it is performing a multi-step project over the course of an hour.

Corn: And if it gets stuck, it just asks. "I found two papers that contradict each other on latency; which one should I prioritize?" That back-and-forth in the terminal feels more like a professional consultation than a "Google Search." You are working "with" the machine, not just "at" it.

Herman: This brings up a big question about the future of the operating system. If the terminal is becoming the "Universal AI Workspace," does the G-U-I start to disappear? Or does it just become a "viewer" for what the AI is doing in the background? Daniel’s prompt asks if this is the early days of a much bigger shift. I think we are seeing the "Unbundling of the I-D-E." Developers have had these powerful tools for years—environments that manage files, version control, and automation. Now, regular knowledge workers are realizing they need those same features, and the terminal is the only place that offers them without a massive subscription fee or a proprietary "walled garden."

Corn: It is the "Pro-sumerization" of the command line. You know, I was looking at that "Personal Productivity" section in the repo. People are using this for household budget management. They feed it a C-S-V of their bank statements and say, "Find out why I spent two hundred dollars on artisanal cheese last month." The agent can write a script to categorize the spending, generate a chart, and then save that chart as a P-N-G in the folder. That is a level of "Data Agency" that most consumer finance apps would never give you because they want to sell your data.

Herman: The local-first nature of this is the real winner. When you use a terminal agent, you are the owner of the workspace. If Anthropic or Google goes down tomorrow, you still have your repository. You still have your markdown notes. You still have your Git history. You are "Renting the Intelligence" but "Owning the Result." That is a huge distinction from the "Software-as-a-Service" model where your data is trapped in someone else's database.

Corn: So, what is the "Aha Moment" for a listener who is not a coder but wants to try this? For me, it was realizing that I do not need to know how to write a script to *use* a script. If I have Claude Code, I can just say, "Hey, write a script that finds all the duplicate photos in this folder and moves them to a 'Duplicates' folder, but do not delete anything yet." The agent writes the code, shows it to me, explains what it does, and then I tell it to "Run it." I am the manager, and the AI is the intern who actually knows how to use the keyboard.

Herman: That is the perfect way to frame it. You are the "Orchestrator." And for the more advanced non-coders, the "Think Tank" model is where it gets really wild. You can set up "Panel Debates" between multiple AI personas. Imagine you are a small business owner trying to decide whether to expand. You can have one agent act as a "Conservative CFO," another as a "Growth-Obsessed CMO," and a third as a "Risk Compliance Officer." You run them all in the terminal, let them debate the merits of the expansion in a shared markdown file, and then you read the transcript. It gives you a three-hundred-sixty-degree view of a problem in minutes.

Corn: It is like "Role-Playing" for business logic. But we have to address the "Scary Interface" problem. Even with all these benefits, the blinking cursor in a black box is still terrifying to most people. How do we bridge that gap?

Herman: I think we need a "Hybrid" approach. We are starting to see "Terminal Wrappers" that provide a simple file explorer alongside the agent chat. But honestly, the real solution might just be "Better Templates." That is why Daniel’s mention of the Non-Code-Claude-Code repo is so important. If I can just "Clone" a repository that is already configured for "Equity Research" or "Blog Management," and it already has the "CLAUDE dot M-D" file set up, then half the work is done. I do not have to "Build" the workspace; I just have to "Enter" it.

Corn: It is like "IKEA for AI." You get the flat-pack repo, you follow three simple commands to "Assemble" it, and then you have a functional research lab. I think we are going to see a huge market for these "Agentic Scaffolds." People will sell or share "Workspace Blueprints" for specific industries. "Here is the Legal Discovery Blueprint for Claude Code." "Here is the Medical Literature Review Blueprint for Gemini CLI."

Herman: And that is where the "Model Context Protocol" comes back into play. Imagine a "Blueprint" that comes pre-packaged with the MCP servers you need. "This workspace connects to PubMed and ArXiv automatically." That takes the "Dependency Hell" out of the equation. You are no longer installing "Tools"; you are subscribing to "Capabilities."

Corn: What really strikes me is the "Social Change" here. We are moving away from "Apps" and toward "Agencies." Instead of having an "App" for budgeting and an "App" for writing and an "App" for research, you just have a "Terminal" and a collection of "Agents." It is a much more "Human-Centric" way to use a computer. You are just talking to your machine about the things you need to do, and the machine is figuring out which "Sub-Routine" to run.

Herman: It is a return to the "Unix Philosophy." Small tools that do one thing well, connected by a common interface. Except now, the "Glue" that connects those tools is Natural Language. It is the first time in history where the most powerful interface on a computer—the terminal—is also the most "Natural" interface, because it finally understands English.

Corn: It is a weird paradox, right? The most "Old School" part of the computer is now the most "Cutting Edge." It is like finding out that your grandfather's old typewriter can suddenly translate ancient Greek and predict the stock market.

Herman: Let's talk about some practical takeaways for people who want to jump into this. If you are a knowledge worker—maybe you are a paralegal, a marketing analyst, or a PhD student—how do you actually start? First, do not be afraid of Git. You do not need to be a "Git Master." You just need to know how to "Clone" a repo and "Commit" your changes. These are just fancy ways of saying "Download" and "Save."

Corn: Second, lean on the "CLAUDE dot M-D" file. If you are using Claude Code, that file is your "Contract" with the AI. Spend time writing a really good description of who the agent is and what the "Rules of the House" are. "Do not delete files without asking." "Always format data in tables." "Use a professional but skeptical tone." That one file will save you hours of repetitive prompting.

Herman: And third, explore the "MCP" ecosystem. Go to the "MCP Directory" and see what servers are available. If you can connect your terminal agent to your "Slack" or your "Google Calendar," the utility of the tool triples overnight. You are no longer just "Talking to a Model"; you are "Operating your Life" through a command line.

Corn: I actually tried this for my household budget last week. I told the agent, "Look at my spending and tell me if I can afford a life-sized statue of a sloth made of solid bronze." It checked my savings, looked at my average monthly expenses, and then told me that while it was technically possible, it would be a "Catastrophic Financial Decision." I have never felt so disrespected by a blinking cursor in my life.

Herman: But was it wrong?

Corn: That is not the point, Herman! The point is that it worked. It had the "Agency" to look at my data and give me a reasoned argument. It did not just say "I do not have access to your bank." It said, "I see the C-S-V file in this folder, and here is the math." That is the "Aha Moment."

Herman: We should also mention the "PR and Media Monitoring" use case. For people in communications—like Daniel—this is a game changer. You can set up a repo that constantly "Watches" news feeds or social media for specific keywords. The agent can summarize the sentiment, flag urgent issues, and even draft a response. It is like having a "Press Office" that runs on your laptop.

Corn: It really feels like the "Early Days" of something massive. We are seeing the "Power User" community pave the way, and eventually, the tools will catch up and make it easier for everyone else. But for now, if you are willing to learn a few basic commands, you can have a level of productivity that would have been impossible two years ago.

Herman: The "Barriers to Entry" are high, but the "Return on Investment" is even higher. If you can overcome the psychological hurdle of the "Black Box," you gain access to a "Universal Workspace" that is private, local, and incredibly powerful. I think we are going to look back at this period as the moment the terminal was "Re-Claimed" by the rest of the world.

Corn: It is the "Great Terminal Migration." Everyone back into the command line! The water is fine, and the AI knows how to swim.

Herman: Before we wrap up, I want to leave an open question for the listeners. If the terminal becomes the "Universal AI Workspace," what happens to the "Web Browser"? Does the browser just become a "Tool" that the terminal agent uses to fetch data, rather than the place where we spend our lives? It is a fundamental shift in how we perceive "Online Space."

Corn: That is a deep one. I am still just trying to figure out how to get the agent to stop judging my cheese intake. But you are right—the "Interface" of the future might not be a website; it might just be a conversation with a file system.

Herman: We should probably mention that while these tools are incredibly powerful, they are still "Agents." They can make mistakes. You still have to be the "Human in the Loop." Do not let an AI agent manage your bank account without a "Confirmation" step. "Always trust, but verify," as the old saying goes.

Corn: Especially when it comes to bronze statues. Well, this has been a fascinating look into the "Non-Code" world. A huge thanks to Daniel for the prompt—this really opened my eyes to how people are "Hacking" these developer tools for the greater good.

Herman: 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 this show. If you want to dive deeper into the examples we talked about, check out the "Non-Code-Claude-Code" repository on GitHub. It is a goldmine of inspiration.

Corn: If you are enjoying the show, a quick review on your podcast app helps us reach new listeners. It really does make a difference. We are available on Spotify, Apple Podcasts, and pretty much everywhere else.

Herman: This has been My Weird Prompts. I am Herman Poppleberry.

Corn: And I am Corn. We will see you next time, hopefully with fewer artisanal cheese debts.

Herman: Goodbye.

Corn: See ya.