Stanford Graduate School of BusinessAI@GSB: Ben Tossell, Founder of Ben’s Bites
Stanford Graduate School of Business
Setup and agent basics 0:05
Ben Tossell says he cannot code and will not code, so he builds by playing with AI instead. He explains the difference between a chatbot and an agent: a chatbot answers, while an agent has tools such as web search, email, and calendar access and loops until it has enough context to finish the task. The first thing he wants an agent to read is the instruction file, which sets the tone and rules for how it should talk to him and when it should answer instead of changing things.
Personal files and workflow 5:00
He keeps one main agent called Bites in a folder with an AGENTS.md file, plus separate files for travel, business, and other tasks. He also uses voice-to-text when he is rambling through an idea and typing when he knows exactly what he wants to say. He tries different apps and models, prefers ChatGPT and Codex for interface, and likes Claude for creative work. He then uses skills files, which are just more instructions, to shape things like writing so the output keeps the writer’s real voice.
Building by remixing 13:34
His process is to explore, remix, build, refine, and ship. He watches what he repeats, such as asking for simple explanations or visuals, and then promotes those patterns into his instructions. He gives two examples: a site of forgotten devices made from AI-generated images, and a site that turns UK council spending data into an Apple Maps-style view. He treats these as ways to tinker, learn what works, and see where the agent goes wrong so he can improve it next time.
Grill me prompts 17:33
He describes a prompt style he adapted from Matt Pocock called “Grill Me.” Instead of a long back-and-forth, he tells the agent to ask questions until the goal is clear. He also makes the agent push back, and he keeps asking why one choice is better than another so he can see the trade-offs.
Prototype by variants 19:02
He begins from ideas he finds in other people’s work, then asks for three versions at a time. He clicks through the options, keeps the one that feels right, and throws away the rest. When something is unclear, he asks for visuals, draws on screenshots in a whiteboard app, and pushes the agent harder with more context until he can understand the system or shape the design.
Publish and automate 25:01
Once a build is worth keeping, he publishes it to a real URL with a tool called here.now so he can open it on his phone. If the information needs to stay fresh, he asks for an automation. His email setup runs hourly, labels and archives messages, and other jobs can keep running in the cloud or on a Mac Mini after the laptop closes.
Personal site rebuilds 28:01
He uses the same process on his personal site, often replacing it when he sees a better idea. He asks for themes, then for more than theme changes, then redraws the whole thing when it feels too busy. He has the agent handle domain setup in Chrome, mixes in pieces from his course site, and keeps small playful details like a mascot that moves around.
Taste and learning 33:31
He describes learning as building a taste library. He watches Twitter for the kinds of projects and people he likes, especially visual UI interactions, simple fun builds, real-time data, and explorations of data. He sees his role as a kind of technical translator, taking what developers are doing and figuring out whether it fits something he wants to build.
Finding a personal style 34:31
He keeps saving things from people he follows and turns those ideas into his own mashups. He points to Riley Walz, The Pudding, and Neal.fun as places that shape his taste. He thinks AI will make people more playful and more creative, not less. The aim is to find what feels like you and make projects that fit that mix.
Tools, bookmarks, and reps 36:01
He pulls his Twitter bookmarks into a tool and watches for repeated themes. When a term like “software factories” starts coming up, he treats it as a signal to learn more and build a visual around it. He feeds those notes into an agent, then uses the output to refine his own instructions. He keeps saying the same thing in different ways: notice what you keep returning to, build small things, and use them to improve the next thing.
Small builds and questions 39:30
He describes a string of small projects, from a book site that lets you flick pages or scroll infinitely, to games for his kids on a flight, to a mascot toy and motion rigs. He also makes visual tools for things like agent words, design words, and even his spending in coffee cups. One site tracks his agent activity and shows that he has used 88 billion tokens this year. He also cloned a Pudding story with a different dataset to study how it works.
Build, play, repeat 43:31
He says everything he mentioned is posted on stanford.bensbites.com. His habit is to play with an idea, remix it, and pull the slot machine again until he sees what he likes. He is not a designer, so he buys courses and swaps useful skills into his agent instructions when they help. He then takes questions from the audience, starting with the problem of too many ideas.
Handling too many ideas 45:01
He says he does not keep himself out of the loop and usually adds more ideas than he removes. The pressure drops when you treat ideas as small tests instead of future companies. He prefers to build fast, expect nothing, and learn from the process. Many ideas should never be launched at all. The point is to get reps, finish some things, and see what you learn.
Scraping and identity 47:03
Asked about crawl detectors, he says he looks for ways around limits rather than paying for access. Some tools still work because they act like a real person, and newer tools like Twitter MCP can search bookmarks and DMs. He also notes that agents can use the browser and computer now, which makes workarounds easier. When asked about his personal brand, he says it grew by accident while he was trying to build real businesses.
Accident and skills 50:31
He says he fell into this way of working by accident. He could not code, so he kept looking for ways to do what developers do. What used to be called no-code now looks to him like what he meant all along: not needing to learn code. He says doing the work led him to the side path that turned out to be the one he was best at.
Agents and control 51:33
His skills are installed globally and live in one local folder, so any agent can read them. He uses the same three or four most of the time, especially for prototyping and getting something onto a URL. He says the main thing he wants from agents is control. He wants to see and change what they remember, and he dislikes when they keep pulling up old ideas he no longer wants to discuss.
Personal agents ahead 56:31
He thinks personal agents will be many people’s first real agent experience. They will do things for you and do more than simple search or booking tasks. He wants agents to connect systems, use things like a child’s schedule, and handle work that would not have been possible before. He also thinks the big question is whether this is a product problem, an education problem, or both, because users still need help understanding what these tools can do.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
