How to Build a Custom GPT, Claude Project, or Gemini Gem in 2025
In the world of AI, the big three – OpenAI’s Custom GPTs, Anthropic’s Claude Projects, and Google’s Gemini Gems – are quick-and-easy ways to interact with large language models (LLMs). Whether you’re a real estate professional looking to build a custom assistant or a founder aiming to scale operations through automation, these tools offer a fast way to create mini AI agents that are tailored, task-specific, and quite useful.
This post breaks down the essential steps to build your own custom GPT, Project, or Gem. By the end, you’ll know how to structure instructions, upload knowledge, refine performance, and test like a pro.
- Interested in mastering AI for commercial real estate? Consider our AI for CRE course and become one of the top 1% of real estate professionals leveraging AI to stay ahead.
What Are These Custom Mini “AI Agents”?
In simple terms, all three platforms let you:
- Define a custom set of instructions (the brain).
- Upload supporting documents or data (the knowledge).
- Enable capabilities/tools like web search or coding.
- Provide prompt examples for users to get started.
Think of them like teaching your junior analyst to perform a very specific task exactly the way you want it to.
Step-by-Step: Building Your Custom GPT, Project, or Gem
1. Start with a Use Case
Clarify the task your AI agent is solving. The narrower and more defined, the better. Examples include:
- A deal screener trained on your investment criteria.
- An email assistant that drafts responses in your tone.
- A leasing bot trained on your company’s policies.
2. Set Custom Instructions
Use a structured prompt framework like RODES:
- Role (R) – Define the AI’s persona. Be specific about who the assistant is and the voice it should adopt.
- Objective (O) – Clearly state the goal. What is the user trying to accomplish with this interaction?
- Details (D) – Provide the blueprint for success. Include requirements, constraints, and context.
- Examples (E) – Include sample inputs and desired outputs. This helps align the AI’s responses with expectations.
- Sense Check (S) – Ask the AI to confirm understanding before proceeding. This minimizes misinterpretation and ensures accuracy.
Here’s how this might look in a commercial real estate context:
- Role: You are a seasoned real estate analyst trained to analyze market trends.
- Objective: Analyze industrial cap rate trends in the Chicago market over the last five years.
- Details: Focus on Class A industrial buildings larger than 200,000 SF from Q1 2018 to Q1 2023.
- Examples: Generate a list of relevant sale comps with accompanying cap rates and sale dates.
- Sense Check: Ask, “Before performing the analysis, do you have any questions?”
3. Upload Supporting Knowledge
This is where your AI gets smart. You can upload:
- Examples of prior work (e.g., past investment memos, deal reviews).
- Process documentation.
- FAQs or client resources.
Ensure your instructions clearly state when and how to use these files. Curate them carefully—they serve as a secondary set of prompts.
4. Add Prompt Starters
Give users a head start. Examples:
- Help me screen this real estate deal.
- Write an LOI for a retail property.
- What’s the cap rate for this investment?
5. Choose Capabilities
Decide what additional capabilities your agent should have, such as:
- Code Interpreter: Essential for working with spreadsheets, PDFs, and data.
- Web Browsing: If your GPT needs real-time data.
- Image Generator: Great for design, branding, or mockups.
- Canvas: Ideal for spatial tools and drawings.
6. Test and Iterate in Preview
Don’t skip this step. Play around:
- Try common user inputs.
- Verify if it uses the right tone and knowledge.
- Check if it stays within scope.
Tweak instructions and re-upload better-curated knowledge if needed.
The edit screen for an example custom GPT.
Pro Tip: Add Signals to Make It Feel Human
Want to make your GPT feel more intuitive? Add a file called user_signals.txt
. In it, define user “signals” like frustration, confusion, or praise and include suggested response styles. This helps the GPT adapt, reassure, or clarify automatically.
Examples:
- Signal: User is confused.
Response: “Totally get that. Want to walk through an example together?” - Signal: User expresses frustration.
Response: “Sounds like a tricky spot. Let’s work through it step by step.”
Naming Matters
Your agent’s name should be:
- Concise and descriptive.
- Consistent with your branding.
- Avoid ending with “GPT” unless it adds value.
Examples:
- Investor Screener
- Tracy Assistant
- CRE Deal Coach
Tools of the Trade
Here’s a simple checklist to ensure a quality build:
- Clear purpose
- RODES-based instructions
- Curated and relevant knowledge files
- Conversation starters
- Appropriate capabilities enabled
- Real-world testing
The prompt screen for an example Gemini Gem.
Closing Thoughts
Custom GPTs, Claude Projects, and Gemini Gems are true amplifiers of your expertise. With a clear purpose, smart instructions, and curated knowledge, your AI agent can become a valuable member of your team.
Frequently Asked Questions about Building a Custom GPT, Claude Project, or Gemini Gem in 2025
What are Custom GPTs, Claude Projects, and Gemini Gems?
These are fast, low-code ways to create personalized AI assistants across OpenAI (GPTs), Anthropic (Claude Projects), and Google (Gemini Gems). They let you define instructions, upload documents, set capabilities, and deploy task-specific AI agents quickly.
What’s the first step in creating one of these agents?
Start with a clearly defined use case. “The narrower and more defined, the better.” Examples include a leasing bot, a deal screener, or an email assistant.
How should instructions be structured for your agent?
Use the RODES framework: Role, Objective, Details, Examples, and Sense Check. This ensures clarity and consistency in how the AI should behave and respond.
What kinds of knowledge files can be uploaded?
You can upload internal documentation, past work samples (like memos or reviews), FAQs, or process guides. These serve as curated data the AI references to stay on task.
What capabilities can be enabled in custom AI agents?
Depending on the platform, you can enable tools like:
Code Interpreter (for spreadsheets, PDFs, and data),
Web Browsing (for live data),
Image Generator, or
Canvas (for drawing/spatial tools).
Why are prompt starters important?
They help users interact with the AI quickly and correctly. Examples include: “Help me screen this deal,” or “Write an LOI for a retail property.” They act as templates for expected usage.
What is a “user_signals.txt” file and why use it?
It defines user moods (like confusion or frustration) and suggests how the AI should respond. For example:
Signal: User is frustrated.
Response: “Sounds like a tricky spot. Let’s work through it step by step.”
What makes a good name for your custom AI agent?
Names should be concise, descriptive, and aligned with your brand. Avoid tacking on “GPT” unless useful. Examples: “CRE Deal Coach,” “Investor Screener,” or “Tracy Assistant.”
What should I do before finalizing my AI agent?
Test thoroughly in Preview. “Try common user inputs. Verify tone. Check if it stays within scope.” Refine your files and instructions as needed.
What is the recommended checklist before publishing?
Use this checklist:
Clear purpose
RODES-based instructions
Curated, relevant knowledge
Useful prompt starters
Correct capabilities enabled
Real-world testing complete