29:07How I AI · Tutorial
Loop engineering for beginners
A plain-English field guide to every loop type — heartbeat, cron, hook, and goal — with two live builds in Claude Code and Codex.
June 17thThe argument in one line.
Winning hackathons repeatedly is a reusable four-step system, not raw coding speed: reverse-engineer the demo that would win first, build it with a stacked AI agent workflow, manage sleep deliberately, and maximize time in front of judges.
Read if. Skip if.
READ IF YOU ARE…
- A student heading into a hackathon who wants a repeatable system instead of improvising on the day.
- A builder already using Claude Code or Codex who wants a sharper agent workflow (skills, MCPs, parallel sub-agents) for any fast-build project, not just hackathons.
- Someone who has shipped hackathon projects before but keeps losing on the pitch, not the build.
SKIP IF…
- You're looking for a line-by-line tool setup tutorial — this is a process and mindset video, not an install guide.
- You have no tolerance for sponsor-read content — a third of the middle section is a dictation-tool pitch woven into the workflow explanation.
The full version, fast.
A five-time university hackathon winner lays out the system behind every win: pick the craziest demo you can think of, check it's feasible in the time limit, then build backward from that one winning moment. Building happens inside a stacked AI agent workflow: dictate prompts by voice instead of typing, run parallel agent workspaces in separate git worktrees, and lean on a small set of skills (an 'interview' skill to lock scope, UI-polish skills, an agitation skill for feeding design feedback, a plain-language skill to stop jargon) plus MCPs for databases and payments. Past the build, winning comes down to managing a sleep deficit on purpose and then maximizing time in front of judges with a live, working demo.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:50
01 · Cold open
Introduces his five-for-five hackathon record and promises the four-step workflow behind it.
00:50 – 02:31
02 · Step 1: Think Backward
Reverse-engineer the exact winning demo moment first, push the idea until it's almost too ambitious, then test it against the 24-hour limit. Uses Twitter and a tool called Cosmos for ongoing inspiration.
02:31 – 04:21
03 · Sponsor: WhisperFlow
Demos the voice-to-text tool he dictates prompts with, showing it clean up a messy spoken prompt into a polished agent instruction.
04:21 – 05:04
04 · Agent harness, explained
Defines 'agent harness' (Claude Code, Codex, Cursor) as the memory/context/routing wrapper around the model, distinct from the agent IDE layered on top.
05:04 – 09:24
05 · Step 2: Agent Workflow
Deep dive on his actual stack: the Superset IDE and git worktrees, his skills (interview, better-UI, animate, agitation, Standard Technical English, Codex sub-agent), MCPs (Supabase, Neon, Stripe), and the Columbia hackathon case study where the system won a 24-hour contest in 8-9 hours.
09:24 – 09:46
06 · Plugin teaser
Mentions he's bundling his whole setup into a one-line installable plugin and asks viewers to comment for it.
09:46 – 10:10
07 · Step 3: Stamina
Treats sleep as a budget to plan in advance for 24-36 hour events, and flags the 4 a.m. energy crash as predictable enough to plan around.
10:10 – 11:39
08 · Step 4: Judging
The pitch strategy: talk to judges as long as possible, lead with motivation and context, then close on a live working demo rather than a faked one.
11:39 – 11:57
09 · Sign-off
Recaps the four steps and closes out.
Lines worth screenshotting.
- Reverse-engineer the winning demo first: picture the exact moment that makes a judge say 'this one's the winner,' then build backward from it.
- Pushing an idea past comfortable and then testing it against the time limit produces more memorable demos than starting from a 'safe' idea.
- An agent harness (Claude Code, Codex, Cursor) is the memory, context, and routing layer wrapped around the model — knowing how it works matters as much as prompting skill.
- Dictating prompts by voice instead of typing produces longer, more context-rich prompts, which measurably improves what the agent builds.
- Running multiple agent workspaces in separate git worktrees lets a team build several features in true parallel instead of one feature at a time.
- A simple 'interview' skill that forces the agent to keep questioning an ambiguous prompt until both sides agree, saved as a spec document, stops hours of work from drifting off the original idea.
- Agents copy an existing design reference far better than they invent an original one, so pointing at inspiration beats describing a vision from scratch.
- More skills and MCPs isn't automatically better: procedural guardrails built to keep weaker models on track can cap what a stronger model would otherwise do unassisted.
- A team won Columbia's 24-hour hackathon in 8-9 hours by spending 2-3 hours on a shared plan, splitting it into independent sub-agent tasks, then doing one hour of manual refinement.
- Plan your sleep budget for a multi-day hackathon in advance instead of deciding in the moment — the energy crash around 4 a.m. is predictable enough to schedule food or shift rotations around.
- In judging, maximizing time in front of the panel does double duty: it deepens their impression of the project and keeps judges away from competing teams.
- A live, working demo beats a polished pitch alone, because judges have learned to discount projects they suspect were faked or weren't actually finished in time.
A repeatable four-step hackathon winning system
BUILD-FAST SYSTEM
Winning repeatedly isn't about typing code faster — it's reverse-engineering the demo first, stacking a deliberate AI agent workflow, budgeting your sleep, and out-talking the judges once the build is done.
02Step 1: Think Backward
- Start from the end: picture the exact demo moment that would make a judge say 'this one's the winner,' then work backward to what you'd need to build it.
- Push the idea until it's almost too ambitious, then test feasibility against the time limit — cut only what genuinely won't fit, not what merely feels bold.
- Track inspiration sources (saved posts, UI feeds) year-round so demo ideas already exist before the hackathon clock starts, not during it.
04Agent harness, explained
- An 'agent harness' (Claude Code, Codex, Cursor) is the wrapper that manages memory, context, and routing around a model — understanding that layer matters as much as prompting skill.
- An agent IDE that runs multiple workspaces in separate git worktrees lets you develop several features in true parallel instead of one linear thread.
- Dictating prompts by voice instead of typing produces longer, more naturally detailed prompts, and that extra context is what improves the agent's output.
05Step 2: Agent Workflow
- A short, repeatable 'interview' skill that forces the agent to question an unclear prompt until both sides agree, then saves that as a spec document, stops a multi-hour build from drifting off the original idea.
- Default to pointing the agent at a concrete design reference rather than asking it to invent a UI from scratch — agents copy far better than they originate.
- MCPs let the agent provision databases and payments directly, but more skills and MCPs isn't automatically better — extra procedural guardrails built for weaker models can cap what a stronger model would otherwise do.
- A team won a 24-hour hackathon in 8-9 hours by spending 2-3 hours on a shared planning document, splitting it into independent sub-agent tasks, then doing one hour of manual refinement.
07Step 3: Stamina
- Treat a 24-36 hour hackathon as a sleep-budget problem, not a stamina problem: plan the sleep deficit in advance instead of deciding in the moment when you're already exhausted.
- The roughly 4 a.m. energy crash is predictable enough to plan around — schedule food or shift rotations with teammates ahead of time instead of pushing through on willpower.
08Step 4: Judging
- In judging, maximize time in front of the panel, not just the quality of the pitch — it deepens their impression of the project and keeps them away from competing teams.
- Lead the pitch with motivation and context before the demo, then close on a live, working demo — judges discount results they suspect were faked or rushed, so working proof beats polish.
Terms worth knowing.
- Agent harness
- The software wrapper around an AI model, such as Claude Code, Codex, or Cursor, that manages memory, context, and tool routing during a coding session.
- Agent IDE
- A development environment built on top of an agent harness that lets you run and monitor several parallel AI coding workspaces at once.
- Git worktree
- A Git feature that checks out multiple copies of the same repository into separate folders simultaneously, so each can be worked on in parallel without switching branches.
- MCP
- A standard that connects an AI agent to external tools and services, such as a database or a payment provider, so the agent can act on them directly.
- Skill (agent skill)
- A saved, reusable instruction set the agent reads to perform a specific task consistently, similar to a cheat sheet the model consults on demand.
- Standard Technical English
- A plain-language rule set the creator built as a skill to stop the AI from explaining things in needlessly technical or confusing jargon.
Things they pointed at.
02:10toolCosmos
02:31toolWhisperFlow
05:05toolSuperset (agent IDE)
05:26tool/interview skill
05:51tool/better-ui skill
06:05tool/animate skill (by Lottie Motion)
06:33tool/agitation skill
06:54tool/STE skill (Standard Technical English)
08:05tool/codex skill
08:44toolSupabase MCP
08:44toolNeon MCP
08:54toolStripe MCP
Lines you could clip.
00:50
“The first and single most important step is to think backward. Imagine you've already won the hackathon.”
02:54
“Nowadays, it's not fast enough for you to just code by hand. If you want to win something, you have to use AI.”
07:20
“Think of skills like a cheat sheet that the agent can read when it needs to do certain tasks to keep it consistent and reliable.”
08:55
“Sometimes having too many skills and MCPs can actually limit the AI by forcing it to adhere to certain procedures... may stop smarter models from reaching their full potential.”
10:10
“At this point, your goal is to yap for as long and as much as possible.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
Hey guys, I'm Jonathan, and over the last year I've gone to five of the most prestigious hackathons in the world at universities like Stanford, Columbia, and Yale, and I've won a prize at each and every single one.
Cloud Code and Codex obviously help me move fast, but it's not as simple as just pasting in a single prompt into your AI agent and going to sleep. So whether you're a student trying to make your first project, or a multiple -time hackathon participant that wants to win, this video will help you.
If cloud has felt dumb to you at times, like explaining things in the most unintelligible way possible or outputting straight AI slop in its designs, then it's likely that you're just missing a critical part of the setup. In this video, I'll walk you through the secret workflow that I use to win against over 1 ,000 other competitors and teach you the four -step process that most other contestants are not aware of to ideate and create actually good projects.
By the end, you too can go and win some hackathons or just create really polished projects that you're proud of. The first and single most important step is to think backward.
Imagine you've already won the hackathon.
Maybe it's a specific track or sponsor that you really like. Now, how did you do it? What crazy demo did the judge see that instantly made them go, yep, this one is the winner?
For me, at TreeHacks 2025, it was a six -foot -tall autonomous food delivery robot. We didn't even shoot for any specific track, but we knew immediately that this project, if completed, would win something. At TreeHacks 2026, it was a three foot tall smart mirror.
At YHack, it was an AI agent that works with you on your computer. And at Columbia, it was the image of four AI agents working in parallel virtual machines on the cloud. The trick I use is to pick the craziest cool thing you can think of and then check if it's possible.
At TreeHacks, I originally wanted to scrape the data of all participants and make a deep fake clone of them. and make those clones go on dates with each other. But I realized that might not be possible in 24 hours.
So keep thinking until you come up with the coolest demo that's still technically feasible. I ended up scraping people's emails to rebuild their closets and find clothing recommendations for my smart mirror instead. And of course, it's important to generally know what's going on in the world.
I like to scroll Twitter a lot for this, and I also find great UI inspiration on Twitter. For example, for Second Self, the project that won first place in personal AI at YHack, I pulled the core idea from a Twitter post that I saw the month before.
One good tool for saving these posts is called Cosmos. It's kind of like Pinterest, but I use it for cool inspiration from all across the web. This is not sponsored, by the way.
Now that you know what to build, the next part is how to build it. Obviously, you need to use an AI agent of some sort to code. Nowadays, it's not fast enough for you to just code by hand.
If you want to win something, you have to use AI. And not only should you just use AI, but even just typing the prompts by hand is too slow. I use a smart voice -to -text tool called WhisperFlow to prompt about four times faster than I can type.
This lets me move so much faster, even though I'm a really fast typist. WhisperFlow isn't just normal dictation. It actively polishes the mistakes.
the restarts, the ums and ahs that happen naturally when you talk. Can you create a feature to basically let users bring in their... And then see how it got rid of all my ums and ahs and it kind of knows what I'm trying to get at even though I stumbled a lot there.
Can you refactor the Barnard CS major so that math UN 2015 can be double counted for linear algebra?
Both math requirements are completed. Yeah. It also formats your prompts by pulling in bullet points or punctuation like you just saw with the numbered list.
I find that talking naturally encourages me to think more freely and this leads to me dictating a lot more context to my agent than if I were to type the same thing. And so this enables my agent to have more context on what I want and then do a better job. WhisperFlow works anywhere that you type, so I use it inside Claude, Cursor, Codex, with their desktop app, and even on my phone with the iOS app.
It's available on iOS and Android. Yo, Alex, you down to go to Butler? Actually, no, let's go to Joe's in like five minutes.
You down?
The good news is that WhisperFlow is completely free to use, and the great news is that the link in my description will get you one month of WhisperFlow Pro with unlimited words for free. Thank you so much to whisper flow for giving you guys this free link and for sponsoring this video It really means a lot and to use an AI agent Well, it's important to know or at least vaguely understand how these agent harnesses work and by agent harness I mean things like cloud code Codex or like cursor the harness is basically a wrapper over the underlying model that handles memories context management Routing etc.
For example cloud code is a harness codex is a harness and cursor agent is also a harness. On top of the harness, you need an agent IDE.
There are a lot out there, but they essentially let you manage many parallel agents at once, which is super important for developing quickly at the hackathon. I use one called Superset. Within Superset, I can create new workspaces, and each workspace lives in its own Git work tree.
which is basically like multiple copies of the code where i can work on a separate feature in each copy and this lets me move a lot faster i like superset because it has its own integrated browser and shows all my code diffs and files in one application so i can really see what the agent is doing for each of these separate features and then on top of that there are skills and mcps that really help too Think of skills like a cheat sheet that the agent can read when it needs to do certain tasks to keep it consistent and reliable.
The most important skill that I use is one called slash interview. This is actually one that I created myself and it's very simple. It just encourages the agent to keep asking me questions on anything that is unclear with my prompt until we are on the exact same page about what to do.
And then it saves that as a document. to refer back to later to prevent deviating from the original vision later down the line. Skills are also how I fixed the problem of Claude speaking unintelligibly and producing slop designs.
Slash better UI is an open source skill that I found that performs a UI audit of your entire app. It recommends interface improvements, animations, and just general polish. Slash animate is a skill by Lottie Motion that applies best practices of motion design and animations to add polished animations into your app.
I use these two skills to add a lot of polish and just the overall feeling of quality to whatever I make at the hackathon. Slash agitation is a secret skill that not very many people know about. It lets you annotate your website and feed those comments along with the ID of the actual element you're referring to.
back into your agent so that the agent has context on what you're talking about. I find that this speeds up my design process by at least three times and it also lets me be a lot more clear and specific about my intentions. Slash STE is a skill I created that instructs Claude to adhere to something called standard technical English.
This is basically like a language rule that just helps reduce a lot of the gibberish and slop that Claude says when it tries to explain something to you. I've noticed that Claude likes to become hyper technical and explain a lot of things that don't really make sense using words like levers, knobs, counterfactual, a lot of stuff that just reads like it makes no sense.
And adhering to standard technical English has really helped me fix this error with Claude. I personally usually start my projects by using the slash interview skill. to write a master plan on exactly what I want to build.
For Opticon, the project that won Columbia's hackathon, I spent two to three hours writing the plan with my friends and having Claude interview me on it relentlessly. After writing a huge document, I had to split the work into independent tasks and then ran one sub -agent for each task. It ran for the next five to six hours, and then I manually refined for one hour with slash agitation.
And then I went to throw my own birthday party and went to sleep, got seven hours of sleep, came back the next morning and won the hackathon. We essentially won a 24 -hour hackathon in only eight to nine hours just by being very thorough with the skills and the planning process. You've also probably noticed that Claude is better at front -end design, whereas Codex is better at back -end.
And since I use both, a nice skill I use is slash Codex. which basically lets Claude spawn in a Codex sub -agent when I want Codex to handle a specific feature or review Claude's work. Agents are also much better at copying.
than at coming up with something completely original. Pointing the agent at an example of a design you like really helps get a good starting point where you can start iterating with the skills we just discussed like slash agentation or the UI ones. MCPs, if you don't know, are a way to connect your AI agent to external tools and services to give the AI more capabilities for the task that it's doing.
For example, letting it manage your database or connecting it to your payment provider. I personally use the Supabase and Neon MCP service to let Claude create databases automatically apply migrations, or create new tables for any feature that I'm building.
I also use the Stripe MCP to set up payments for any apps that I like to work on. In general though, agents are getting much, much better very quickly. And so sometimes having too many skills and MCPs can actually limit the AI by forcing it to adhere to certain procedures that were needed in the past to keep dumber models on track, but may stop smarter models from reaching their full potential.
I recommend typing slash skills into your cloud terminal and turning off all skills that aren't critical. Cloud also has a system called plugins, which basically lets people bundle a bunch of skills and MCPs into one thing that you can install really quickly. If you're curious about my workflow in particular, I'm working on a plugin myself to bundle everything that I use into one thing that you can install in just one line.
So comment plug -in if you want it and I'll send it to you. Now that you have a great idea, great implementation, your agents are cooking, the next step is to not sleep. Those hackathons are 24 to 36 hours long.
And honestly, if you want to win, you can't sleep that much. For 24 hour hackathons, I tend to just pull an all -nighter. And for 36 -hour hackathons, I like to sleep five hours in the first night and then don't sleep the second night.
Red Bull and anything caffeinated is really helpful for this. But for me personally, I low -key don't even need caffeine because vibe cutting just gets me going. Honestly though, it gets really bad around the 4 a .m.
mark, so I recommend eating something around then or taking shifts with your teammates. The next and final step is judging. At this point, you've come up with a killer idea.
and you've spent the last 24 to 36 hours grinding on your project. Now, all that's left for you to do is pitch it to the judges. At this point, your goal is to yap for as long and as much as possible.
You want to talk to the judges for as long as possible. Not only does this make them like your project more and give you more time to explain all the technical intricacies and all the things that you had to consider while building your project, but it also keeps the judges away from competitors. This might seem kind of cutthroat, but that's just how the competition goes.
At this point, your amazing project idea and your beautiful technical implementation should just be enough to just blow the judges away and you'll see them get visibly engaged as you give your demo. At this point, it's up to you to really highlight all the technical intricacies of what you did and all the challenges you had to overcome.
That way you'll leave a lasting impression on the judges. At YHACK, I talked in depth about my motivations for building Second Self, the project that I worked on, the state of the world of AI and where I thought it was heading and why I built this project. and then most of it on a live demo of the product itself.
Live demos are particularly impressive at hackathons because people tend to fake a lot of stuff that they can't get done in time. So showing a live demo that actually works will show them how much effort you put into the project and the fact that the project actually works. Finally, you'll probably fall asleep around the time the awards ceremony is going on from just the lack of sleep over the last two days, but hopefully you'll wake up to your name on the big screen and a nice prize in your pocket.
Anyways, those have been my top tips for how to win a hackathon and the exact four -step workflow that helped me win five out of five hackathons in the last year. Let me know what you think and if it was helpful and I'll see you guys in the next one. Peace.
The Hook
The bait, then the rug-pull.
He's won a prize at five straight university hackathons against fields of a thousand-plus competitors, and he says it isn't the code that wins it. It's a four-step system he runs every time, from the idea to the judging room.
Named ideas worth stealing.
00:50list
The Four-Step Hackathon Workflow
- Think Backward
- Agent Workflow
- Stamina
- Judging
The creator's repeatable process for every hackathon: pick the winning demo first, build it with a stacked AI agent workflow, manage sleep deliberately, then maximize pitch time with judges.
Steal forany time-boxed build sprint — an internal hackathon, a demo day, or a weekend MVP push
04:21concept
Agent Harness vs. Agent IDE vs. Skills vs. MCPs
- Agent harness (Claude Code / Codex / Cursor) — manages memory, context, and routing
- Agent IDE (e.g. Superset) — runs parallel workspaces in separate git worktrees
- Skills — saved instruction sets the agent reads for consistency
- MCPs — connect the agent to external tools like databases and payment providers
A layered mental model for how a modern AI coding stack is actually built, from the model wrapper up to task-specific tooling.
Steal forexplaining or auditing any AI-agent coding setup
How they asked for the click.
VERBAL ASK
03:52product
“the link in my description will get you one month of WhisperFlow Pro with unlimited words for free”
Woven into a live demo of the tool mid-video rather than a separate ad break, so the pitch and the proof land together.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
