digest
Short Summary
Main recommendation: Build a custom GPT in OpenAI's GPTs builder by combining a knowledge base (uploaded training documents) with a custom action that calls an external API, using a Replit-hosted solar calculator as the example. The speaker demonstrates this end to end with a fictional company, "Smith's Solar," and says the whole setup takes about five minutes.
Key Points
- The knowledge base handles informational questions, such as sales offers and CRM update procedures, using uploaded company training documents.
- The custom action lets the GPT perform a task: the sales rep gathers client data (power bill and roof size), the GPT sends it to an API, and the API returns estimation data to share with the client.
- The solar calculator back end runs on Replit, so no separate server setup is needed beyond forking and running the template.
- The speaker tests both functions in the GPT chat, confirming the knowledge answer (early bird discount) and the calculation request.
Details Worth Noting
- Concrete instructions:
- Click Create on the ChatGPT homepage, then open the Configure section.
- Enter a name (the example uses "Smith's Solar Sales Assistant") and a short description.
- Paste a detailed instructions prompt that defines the assistant's role, says it should send roof size to the solar calculator action, and says it should keep CRM records updated.
- Upload the training document in the Knowledge section (the example uses "Solar Sales Rep Training 1").
- Fork the Replit template first (click Run and open the New Tab to get it running), then go back to the GPT Builder's "Add actions" Schema area and paste the provided code snippet describing the functions, inputs, and outputs.
- Replace the URL in the schema with the URL of your running Replit page, removing the trailing backslash, so it reads the schema and adds "get and calculate."
- Optionally generate an image with DALL·E; the speaker skips conversation starters.
- Test with a knowledge question and a calculation prompt, such as a 12-meter-wide roof and $450 monthly power spend.
- Save the GPT and choose "Only me" to keep it private, then confirm.
- The company, "Smith's Solar," is explicitly fictional and created only for the tutorial.
Bottom Line
A custom GPT becomes useful when it pairs company documents for answering questions with an external tool for doing calculations, and the speaker presents this as a quick build that others can copy from his shared template.
