In his final quarter at UChicago, fourth-year Ben Heim has been taking his own advice. As president of the International Leadership Council (ILC) Startup RSO, Heim mentors students on how to build startups. When students feel stuck, he offers them a simple piece of wisdom: put your head down and start building, even if the project is small. “I’m a big believer that ideas come after doing, not before,” he said.

Each week, Heim, a computer science major, follows this advice by building a new project using Claude, a large language model (LLM) developed by Anthropic, documenting his progress on his personal website along the way. He sustains this productivity by writing down one problem he is curious about each day; then, at the end of the week, he selects one to solve with the help of Claude Code.

Six weeks in, Heim’s projects have aimed to address a wide range of both personal problems and community needs. So far, he has built an airport assistant, a workout tracker, and a coffee shop app, as well as several more expansive projects.

A Personal AI Assistant

Heim’s first project took shape when he needed to submit a purchase request for a club. “It’s a simple thing to do,” he said, “but I thought, ‘could I get an agent to just fill out the form for me?’” He wanted to use LLMs as employees rather than thought partners so that he could delegate mundane tasks without the need for back-and-forth interaction.

The concept of the project is similar to OpenClaw, an AI agent that can execute tasks on a user’s computer. However, because he wasn’t comfortable giving a machine full access and control over his computer, Heim instead resolved to build a similar agent of his own.

To address safety concerns, he set up a Virtual Private Server (VPS), a remote, private portion of a larger physical computer, allowing him to isolate and safely monitor the agent’s activity. Then, using the Claude Agent Software Development Kit (SDK), a toolkit that allows developers to build AI agents, he constructed a system through which the AI could execute actions—including submitting purchase requests—within this server on his behalf. The Anthropic Application Programming Interface (API) enables the agent to interact with outside applications, so Heim’s agent is synced with the to-do list app he uses, Todoist, and interacts with the user through a Telegram chat interface.

The agent can be activated either by giving it the command word, “go,” which signals that it should begin executing certain actions, or by tagging tasks in Todoist with the label “delegate,” which the agent then detects and takes over. Thus far, Heim says the system has proven capable of handling both cognitive and procedural tasks, from conducting research to booking flights. It also texts the user updates on its progress, sometimes requesting input where necessary, such as login or credit card information.

An Airport Assistant, Gym Tracker, and Coffee Shop App

Some of Heim’s projects have come from specific personal needs, including the desire to calm his mother’s recurring anxieties about whether he will arrive at the airport on time. His second project was an airport assistant app that tells the user when to leave for the airport by factoring in the time it will take to get there, the TSA wait time, and desired buffer time. The tool syncs with the user’s Google Calendar to scan for upcoming flights and then, on the day of the flight, texts the user when they should leave for the airport. This was done “more for my mom than for me, honestly,” Heim said.

Heim has also drawn inspiration from hobbies. An avid weightlifter, he relied for a long time on spreadsheets to track his workouts over time but felt there was something missing. “The thing about lifting is that progressive overload is really important,” Heim said. “You want to be lifting a little more than you did last time. I just wanted something that would tell me: am I actually doing that?”

A week later, he built a workout-tracking app, which logs the user’s lifts and compares them to past workouts, nudging them to do a little more if they’re lagging behind their previous attempt. The interface is color-coded, with red indicating that the user is doing worse than their last workout and green indicating that they are improving.

Heim already uses the app consistently with his girlfriend but hopes to reach more users by releasing it on the iOS App Store. “The broader philosophy here is that there are a ton of workout apps out there, but with vibe coding you can just build something custom for yourself and maintain it pretty cheaply,” Heim said. “Why not make exactly what you want?”

Heim’s work has also extended beyond personal use, leading to collaborations with local businesses. Like many UChicago students, Heim is a frequent coffee drinker who often relies on mobile-ordering through apps of major chains like Starbucks and Dunkin’ to order his coffee in advance and pick it up on his way to campus.

“For independent coffee shops,” Heim explained, “that kind of thing doesn’t really exist.”

His third project aimed to extend this technology to local coffee shops. The app, which is expected to be rolled out “in the next few weeks,” is designed for Robust Coffee Shop in Woodlawn, and will allow its customers to order online, pick up their coffee, and access digital rewards. Heim hopes that this will lead to further collaboration with other coffee shops.

Of course, not every project turns out equally successful. “Last week, I built something that was honestly pretty bad,” Heim said.

Heim’s most recent project grew out of an interest in context engineering, the process of organizing and optimizing the information provided to an LLM. “A big part of what makes AI agents good or bad is the context you give them,” Heim said. “So I wanted to visualize what context my agent was actually pulling in.”

The visualizer displayed the agent as a central node—an individual point in a visual map—and represented the pieces of context it pulled in as connected nodes and edges—the lines connecting those nodes—representing relationships between items in a graph. The goal was to visualize how information was being used during the AI’s execution.

Though the visualizer ultimately proved largely unhelpful and dysfunctional, Heim emphasized the importance of the process itself. “I realized that you can’t shoot for the moon with AI automation and context engineering,” Heim wrote in last week’s entry on his website. “You need to build small building blocks that work effectively and then go from there.”

Heim has since moved away from personal tools and is currently collaborating with a friend to construct systems that can address larger structural problems in institutional data management. Many institutions rely on vast amounts of stored data, much of which becomes contradictory or outdated over time, he explained. “The idea is to use AI agents to continuously prune and clean that data, so that when you go to query it, you’re actually getting accurate, up-to-date information,” he said.

“If I Didn’t Have Claude, None of These Projects Would Ever Exist”

Many associate AI with ease and convenience. Although the end goal of AI tools is, in many cases, to optimize efficiency, the process of building those tools is itself very complicated, demanding a comprehensive understanding of how LLMs function.

Heim emphasized that most of his projects could, in theory, have been built without AI, but the efficiency offered by Claude makes the development process more practical and accessible. “If I didn’t have Claude, none of these projects would ever exist,” Heim said. “They’d just take too much time and energy to be worth it.”

Despite being technically skilled as a programmer, Heim rarely interacts directly with the code of his projects, which are mostly written using a mix of Python and JavaScript. “Because Claude Code handles it all, I can just think about high-level infrastructure questions and don’t really need to think about the codebase,” he said.

The most significant component of the construction of these apps, therefore, is prompting. Claude Code comes with a Claude.md file, a document where the user can write instructions that automatically show up every time they prompt the AI. In this ‘quasi-markdown file,’ Heim includes two key prompts, which are fed into the AI each time he embarks on a new task.

First, to prevent hallucinations—confident, but false, responses generated by AI models—Heim has the model rate its confidence on every response. “That alone helps a lot, because it stops the model from being overconfident or just appeasing me,” he said.

Second, he offers the agent a mechanism through which to verify its own work, explain how it verifies it, and give him the results. “I think of myself as a non-technical CEO with a really capable software engineer underneath me—who is also sometimes incredibly dumb, so you have to figure out how to work with them,” Heim said. “It’s better than me at coding, and I’m a CS major, so that’s saying something.”

Heim also uses AI to generate ideas for future projects. While many of Heim’s ideas originate from personal needs, such as the airport assistant, workout tracker, or coffee shop apps, Heim uses an AI agent—which he built himself—to assist in coming up with ideas. “[It] looks at the problems I’ve been identifying and the writing I do every day and tries to find connections and suggest things that would be useful for me to build,” Heim said.

The Future of AI: “Most People Aren’t Thinking About It Enough”

For now, Heim’s work largely centers on practicality and efficiency: helping to build tools faster and automating tasks. In the future, however, he believes that AI may develop the capacity to innovate in ways that humans cannot currently fathom. “If I could say the full capabilities of AI with enough confidence, I’d be a trillionaire,” Heim said.

Heim pointed to the rapid growth of models over the last few years, suggesting that many people underestimate its abilities and potential. While challenges persist for LLMs, such as hallucination, advances in machine learning have allowed AI to expand its capabilities to the extent that they can perform complex tasks such as diagnosing medical conditions, powering autonomous vehicles, analyzing large datasets, and generating human-like text and realistic images.

It is uncertain where AI innovation will lead us, but for Heim, it has the potential to revolutionize the world. “There’s this J-shaped productivity curve where things actually get a little worse before they suddenly skyrocket, and I think we might be right on that brink, which is really exciting but also really scary,” he said.

Heim expressed frustration at the lack of discourse surrounding the direction of AI, arguing that people across industries should be thinking more critically about the long-term consequences of AI implementation. “Most people aren’t thinking about it enough—even just the basic question of ‘how can we use this to actually help humans?’” he said.

That urgency extends to AI safety, a field concerned with ensuring that AI systems are secure and aligned with human values, which Heim considers a very important and pressing issue. “If you’re someone who has a super negative emotional reaction to AI, you’re the person that should be exploring it and using it most, to figure out what’s actually going on and how we can use it to help people,” Heim said.

Achieving that level of understanding requires going beyond surface-level interaction and obtaining a more substantive grasp of the functions and capabilities of AI. Most people’s experiences with AI begin and end with typing questions into a chatbot, but, according to Heim, that barely scratches the surface of what these tools can do.

“If you’re still just chatting with Claude or ChatGPT in the browser interface—that was [the] frontier two years ago—that’s nowhere near what these things can do now,” he said. “There’s no class for this, so you just have to figure it out on your own.”