Introduction
One year ago I challenged myself to go build a startup, from "no idea" to exit, in just one hour using ChatGPT - and I did it! Now that we have Autonomous AI Agents at our disposal, can we use them to build a startup in just one minute? Let's see...
The One Hour Startup
Immediately after ChatGPT was released last November, I started using it to automate various tasks that are typical in a technology startup: Creating investor pitches, recruiting employees, crafting marketing content, producing product specs, and much more. A few weeks later I decided to challenge myself: Would it be possible to go through the full lifecycle of creating a technology startup, from the initial idea to the exit, using ChatGPT alone, and do all this in just one hour? 60 minutes later, and the mission was completed!
I was really impressed with the results, which were achieved using GPT-3.5 (GPT-4 was not publicly available at that time). I posted this on LinkedIn, and the post quickly went viral with over 140 shares and 600,00 organic views. The purpose of this experiment wasn't to build a startup that has an actual product and customers, but rather to show entrepreneurs and company executives that ChatGPT can be used to accelerate every function in the company. I "open-sourced" the 68-page transcript of my conversation with ChatGPT to serve as a toolbox for startup teams, showing them where and how ChatGPT can be used to supercharge their operations.
So what did I cover in that project? I started by pretending I have no idea about startups, and asked ChatGPT what would be the stages of creating a startup.
Pretty much makes sense, right? Except that I would probably try to do some fundraising before building the MVP… Then I asked ChatGPT to go through this whole process, starting from step one.
I didn’t like any of them, so I asked for more.
From this list I selected number 6, and off we go!
I chose No. 10, and the next step was a slogan.
Not bad, number 2 was my personal favorite.
Enough with branding, let’s get down to business!
This makes sense… After diving into the details of some of these models, I asked for an outline of an investor presentation, an org chart for the company, and a job description for the position of the CTO.
I then asked ChatGPT to analyze the LinkedIn profile of a potential CTO, comment on his suitability, and prepare interview questions.
Then I moved on to more technical topics, including writing a spec for the product, a detailed spec for one of the functions, and even python code for implementing that function.
Next step - some marketing content…
And finally - I asked ChatGPT who would be potential M&A candidates once the company reaches maturity. I got a pretty interesting list, and then asked it to write the M&A contract with that company. THE END - From no idea to exit in 60 minutes!
The One Minute Startup
I was recently invited to present my “One Hour Startup” project at the Product For Product (P4P) community sponsored by Monday.com. Since the original project took place last January, I decided to challenge myself further: Can I reach the same outcome, of creating all the documents needed for a startup from no idea to exit, in just one minute? I tried several different approaches, as outlined below…
Method 1: One Large ChatGPT Prompt
In the original One Hour Startup project, I gave ChatGPT separate prompts for each resource I needed, starting from the idea and the branding, and then going through the required documents for HR, Product, Marketing, etc. So my first thought was: What if I write one big prompt that asks for the startup idea AND all of these documents at once? This is the prompt I used:
"Suggest an idea for a startup that is innovative, based on deep technology, and can bring significant revenues. Create all the resources and documents for this startup including name, slogan, business model, pricing tiers, channels to market, competitive analysis, investor presentation, investor cover letter, org chart, job descriptions and interview questions for senior management, office lease agreement, product specification, detailed specification for a single feature of the product, python code for part of that feature, development plan, marketing plan, website structure, website content, blog posts, Facebook posts and Linked posts, list of potential acquirers, and M&A contract with one of the potential acquirers."
You can see ChatGPT’s full response to this prompt here.
Basically what happened was that ChatGPT gave a very concise version of those documents: An outline, some section names, a few ideas etc., and prompted me to ask for additional details on each if needed.
It started out quite promising, giving me a startup idea, name, slogan and 3 different business models.
Then, it followed up with potential channels to market (headlines only), competitors (name only, no analysis), and a very thin outline of an investor presentation (slide names only).
All the plans were also very slim, the website structure included only the main menu, and instead of home page content it provided some random placeholders.
Finally, instead of an actual blog post I only got the title, instead of social media posts I got their topics, and the M&A contract was a 5 word placeholder.
So what happened? My initial guess is that I’ve hit the limit of ChatGPT’s context window, which was 8000 tokens at the time (before the introduction of GPT-4 Turbo). But the length of the response is less than 600 words… So it may have been just the built-in “laziness” of ChatGPT - it sometimes tries to do the minimum work possible, and lets you “enjoy” doing the bulk of the work… If I “push” ChatGPT by asking (in a single follow-on prompt) for the full content of all documents I get some more, but much less than the content I got in my original project. If I request to elaborate on each document in a separate prompt, I get content that is similar to my original outcome, but then again that would take a full hour…
Method 2: AgentGPT
After the failure of the “single large prompt” method, I decided to experiment with autonomous AI agents. Autonomous agents are software entities that receive a prompt for a high-level task, then break down the task to smaller sub-tasks, launch separate AI chatbots to perform those sub-tasks, and finally integrate the information from all the sub-tasks into a single response. Autonomous agents can also use other tools such as image and speech generation, web access, storage and code execution to complete the sub-tasks. Most of the implementations are open source, with some notable examples being Auto-GPT, Jarvis, BabyAGI and Personoids. I selected an agent called AgentGPT, which has a convenient web interface at agentgpt.reworkd.ai.
This time I used a much simpler prompt:
“Suggest an idea for a startup that is innovative, based on deep technology, and can bring significant revenues. Create all the resources and documents for this startup.”
Since AgentGPT can break down the task into smaller sub-tasks, I didn’t have to spell out all the required resources - I expected AgentGPT to discover them on its own, and then launch each one as a sub-task. And this is exactly what happened!
AgentGPT took my request, and broke it down to sub-tasks: It decided that first it needs to identify deep technologies that have potential for innovation, then it should brainstorm startup ideas that leverage those technologies, evaluate the feasibility and market potential of each idea, and then create all the resources for the selected startup idea. In the last task, it specified that those resources should include a business plan, marketing strategy and financial projections (I wanted it to also create presentations, specs, marketing material etc. but still that’s a great start).
The nice thing about autonomous agents is that they are… autonomous! So once you submit the prompt for your high-level task, you can grab some popcorn and watch your screen in amazement as the agent does its thing. No further input is needed! You can see the full output of AgentGPT for my prompt here.
So here is AgentGPT identifying those deep technologies with strong potential…
Interestingly enough, AgentGPT also identified the AI Healthcare segment as the best one to pursue for an innovative, revenue-driven startup - same as ChatGPT in my previous single-prompt example. Then it listed all the required resources for a startup in this field, this time adding more resources such as technology roadmap, prototype development and team building.
Then AgentGPT executed the second task - finding a specific startup idea within this field.
A Virtual Reality training platform for healthcare professionals! Why didn’t I think of that… And again, for some reason AgentGPT is “reminding” itself of the required resources for the startup, this time adding technology development, partnerships and collaborations, and user Interface & experience design.
Then the agent suggested another idea: An AI platform for personalized medicine.
For this idea, it analyzed Feasibility, Market Potential and Revenue Generation Channels, and then again listed the required resources, this time including business plan, marketing plan, financial projections, product development and legal documents.
Now it finally reached the stage of creating actual resources for the first idea (VR platform for healthcare) - business plan, marketing strategy and financial projection. As you can see it created only outlines of the first two documents, and one bullet for the third. But then came a surprise…
AgentGPT launched a new task for creating a financial model for the startup, and then executed that task by writing Python code! The numbers are just placeholders, but that’s a very nice effort.
I guess this is where my free plan tokens for AgentGPT expired, so it created a summary of the work done so far and exited.
My take on this: With the autonomous agent approach you get a much better result than with the “ChatGPT large single prompt” approach. The agent creates tasks based on your requests, executes them one by one, creates additional tasks as needed, and then provides a nice summary at the end.
Now, what if ChatGPT itself could be persuaded to act as an autonomous agent? We’ll see that in the next section…
Method 3: ChatGPT Agent Prompt
Alexander Lerivag crafted a mega-prompt that turns ChatGPT into an autonomous agent. It does this by asking ChatGPT to alternate between the roles of several different advisers, and a project manager. The advisers are chosen according to the task specified by the user.
So I took this mega-prompt, and inserted the same goal I gave AgentGPT:
“Suggest an idea for a startup that is innovative, based on deep technology, and can bring significant revenues. Create all the resources and documents for this startup”.
In response to this prompt, ChatGPT immediately created 4 advisers: Technology and Innovation, Market Research and Analysis, Business Modeling and Strategy and Legal and Compliance. Then the advisers started to work on building the startup, and all I had to do was press “n” to confirm that the team could proceed. You can see the full ChatGPT session here.
The results were better than the “ChatGPT large single prompt” method: The advisers came up with some potentially interesting fields, I selected one of them, and then the advisers proceeded with creating a full “startup plan” which included value proposition, partners, channels, market size, competitors etc. But…they didn’t come up with a specific startup idea! So this plan was very general, and could fit any startup within the selected field… After presenting the plan, I asked to add more detail, but the added details didn’t have any “meat” since again, there was no specific idea to describe.
I noticed that I couldn’t get results using this method, so I tried another approach: I asked the team to create all the required documents for the startup, this time giving all the document titles. My Project Manager immediately created a new team of advisers, this time consisting of Business Modeling and Strategy, Market Research and Analysis, Legal and Compliance, Product Development, Human Resources and Marketing.
The Project Manager warned me that the task is huge so I’ll only get outlines, and I approved. Indeed I got outlines of the document at a very high level, similar to what I got when using the same prompt without a team of advisers (after all, it is the same ChatGPT…). Then the Project Manager asked if it should expand the document. I confirmed, and got some more detail on each section. Then it proposed to expand again, I confirmed again, and more details emerged, but still the documents were very thin.
Finally, the Project Manager offered to send me the comprehensive package of all the detailed documents. I agreed of course, but the package never arrived, and then the Project Manager claimed that the documents can’t actually be sent due to the limitation of the system (at that time, ChatGPT couldn’t output files). So I didn’t have any choice, and confirmed receiving the full package “virtually”...
Conclusion
With the "One Minute Startup" project I tried to see whether AI chatbots and autonomous agents can create a startup with all its resources using a single prompt. While it's clear that current AI tools cannot achieve this goal, the experiment highlighted their value for assisting entrepreneurs and business executives in many tasks.
Each method had its strengths and weaknesses. The single large prompt was quick but lacked depth, AgentGPT offered more detail but tended to diverge when working unattended, and using ChatGPT as a group of advisers was fun but also didn’t reach the desired results.
This experiment underscores the importance of human oversight and the complementary relationship between AI and human creativity and judgment in the business world. Instead of “send and forget”, with current technology we still need to keep “an eye on AI”, steering it in the right direction and supervising its actions. But even when we humans slow down AI, the combined power of human+AI still provides an unprecedented productivity boost.