Last updated: February 2026
How to Create Custom AI Agents:GPTs, Gemini Gems, and Claude Projects
Custom AI agents let you build specialized chatbots without any coding. Custom GPTs (ChatGPT) offer the most features, including API connections, image generation, and a public store. Gemini Gems (Google) excel at Google Workspace integration. Claude Projects (Anthropic) provide the best writing quality and largest knowledge base capacity. All three require a paid plan at approximately $20 per month. Choose based on your ecosystem and primary use case.
The definitive guide to building, configuring, and deploying custom AI agents across all three major platforms. Includes step-by-step tutorials, prompt templates, and a head-to-head comparison.
What Are Custom AI Agents?
Custom AI agents are personalized versions of large language models that you configure to behave in specific ways. Instead of getting a generic chatbot that tries to do everything, you create a focused assistant that knows your business, follows your rules, and speaks in your voice.
Think of it this way: a standard AI chatbot is a general-purpose intern on their first day. A custom AI agent is a trained specialist who has read your handbook, knows your processes, and understands exactly how you want things done.
All three major AI platforms now offer their own version of this concept:
Custom GPTs
OpenAI / ChatGPT
The most mature platform. Build chatbots with custom instructions, knowledge files, API actions, and DALL-E image generation. Share publicly through the GPT Store.
Gemini Gems
Deep Google Workspace integration. Create AI personas that work natively with Docs, Sheets, Gmail, and Calendar. Ideal for teams already in the Google ecosystem.
Claude Projects
Anthropic
Best for writing and analysis. Upload large knowledge bases, set detailed style preferences, and get the most nuanced, well-reasoned responses of any platform.
Below, I break down each platform in detail: how to create an agent step by step, what works, what does not, and which one is right for your specific needs.
Custom GPTs (OpenAI / ChatGPT): The Complete Guide
Custom GPTs are the most feature-rich custom AI agents available. Launched in November 2023 and significantly expanded since, they let you create specialized chatbots that combine custom instructions, uploaded knowledge, API integrations, web browsing, image generation, and code execution in a single agent.
What Are Custom GPTs?
A custom GPT is a tailored version of ChatGPT that follows your specific instructions, has access to your uploaded documents, and can optionally connect to external services through API actions. Once created, it lives in your ChatGPT sidebar and behaves exactly as you configured it every time you open a conversation.
Unlike a standard ChatGPT conversation where you paste your instructions every time, a custom GPT remembers its role, rules, and knowledge permanently. It is the difference between briefing a contractor before every task and hiring a full-time employee who already knows the playbook.
How to Create a Custom GPT: Step by Step
Open GPT Builder
Go to chat.openai.com and click "Explore GPTs" in the sidebar, then click "Create" in the top right. You need a ChatGPT Plus subscription ($20 per month) or higher. The GPT Builder interface has two tabs: Create (conversational setup) and Configure (manual setup). I recommend starting in the Configure tab for more control.
Set the Name and Description
Give your GPT a clear, descriptive name that signals its purpose. The description appears when others discover your GPT, so make it specific. Example: instead of "Marketing Helper," use "B2B SaaS Email Campaign Writer" so users immediately know what it does and whether it matches their needs.
Write Your Instructions
This is the most important step. Your instructions define the GPT's personality, behavior, knowledge scope, and rules. The Instructions field accepts up to about 8,000 characters. Structure your instructions with clear sections:
- Role definition: Who is this agent? What is its expertise?
- Behavioral rules: How should it respond? What should it avoid?
- Output format: Should it use bullet points, tables, specific lengths?
- Edge cases: What should it do when it does not know something?
- Conversation starters: What questions should it ask first?
Upload Knowledge Files
Upload up to 20 files that your GPT can reference during conversations. Supported formats include PDF, DOCX, TXT, CSV, JSON, and more. This is how you give your GPT domain-specific knowledge: product documentation, FAQs, style guides, process documents, pricing sheets, or research papers. The GPT uses retrieval to search these files when relevant to a user's question.
Configure Capabilities
Toggle the built-in capabilities your GPT needs:
- Web Browsing: Lets the GPT search the internet for current information
- DALL-E Image Generation: Lets the GPT create images from text descriptions
- Code Interpreter: Lets the GPT write and execute Python code, analyze data files, and create charts
Only enable what your GPT actually needs. A customer support bot probably does not need image generation. A data analysis agent definitely needs Code Interpreter.
Add Custom Actions (Optional)
Actions connect your GPT to external APIs. You define an OpenAPI specification (a JSON or YAML file describing API endpoints), and your GPT can call those APIs during conversations. This is how you build agents that check order statuses in Shopify, create tickets in Jira, look up contacts in HubSpot, or query your own database. Actions are the most powerful feature of custom GPTs but require some technical knowledge to set up.
Test and Publish
Use the preview panel on the right side of GPT Builder to test your agent in real time. Ask it questions that test edge cases, not just the happy path. When satisfied, click "Save" and choose your sharing setting: Only me (private), Anyone with a link (semi-public), or Everyone (listed in the GPT Store).
Best Practices for Custom GPT Instructions
- Be specific about what the GPT should NOT do. Constraints are more reliable than open-ended directives. "Never provide medical diagnoses" is clearer than "be careful with health topics."
- Include example interactions. Show the GPT a sample user question and your ideal response. This grounds the behavior more effectively than abstract rules.
- Define the first message. Use Conversation Starters to guide users into providing the information your GPT needs to help them.
- Structure instructions with markdown headers. The model processes structured instructions more reliably than wall-of-text paragraphs.
- Iterate based on real use. Your first version will have gaps. Use it for a week, note where it fails, and refine the instructions.
Example Use Cases for Custom GPTs
Customer Support Bot
Upload your FAQ database and product docs. Add API actions to check order status. The GPT handles tier-1 support questions 24/7, escalating complex issues to humans.
Writing Assistant
Upload your brand voice guide and sample content. The GPT writes blog posts, emails, and social media content that matches your established tone and style.
Research Analyst
Enable web browsing and Code Interpreter. Upload your research framework. The GPT searches the web, analyzes data, creates visualizations, and produces structured reports.
Course Companion
Upload course materials and lesson plans. Students interact with the GPT to get explanations, quiz themselves, and work through exercises at their own pace.
Limitations of Custom GPTs
- -Requires ChatGPT Plus ($20/mo). Free-tier users cannot create or use custom GPTs.
- -No free API access. Using custom GPTs through the API requires separate API billing, which can be expensive at scale.
- -Knowledge file retrieval is imperfect. The GPT does not always find the right information in uploaded files, especially with large or complex documents.
- -Instructions can be jailbroken. Determined users can sometimes get the GPT to reveal its system instructions. Do not put sensitive information in the instructions field.
- -Limited analytics. You get basic usage counts but no detailed analytics on how users interact with your GPT.
- -GPT Store discovery is limited. Publishing to the GPT Store does not guarantee users will find your GPT. Discoverability depends on SEO, reviews, and category ranking.
Gemini Gems (Google): The Complete Guide
Gemini Gems are Google's answer to custom AI agents. Introduced in 2024 under the Google AI Pro plan, Gems let you create custom AI personas with specific instructions that persist across conversations. Where Gems really stand out is their native integration with Google Workspace, making them the clear choice for anyone whose workflow lives in Google's ecosystem.
What Are Gemini Gems?
A Gem is a custom version of Gemini that follows your instructions every time you start a conversation with it. You define the Gem's personality, expertise, response style, and rules. Unlike a regular Gemini chat where you paste your context each time, a Gem retains its configuration permanently and is accessible from your Gem library with one click.
Gems also benefit from Gemini's massive context window (over 1 million tokens), which means they can process extremely long documents, entire codebases, or lengthy conversation histories without losing track of earlier context.
How to Create a Gemini Gem: Step by Step
Open Gem Manager
Go to gemini.google.com and look for the "Gem manager" option in the left sidebar. You need a Google AI Pro subscription (see one.google.com/about/plans for current pricing). This plan includes Gemini features across Google Workspace apps. Click "New Gem" to start building.
Name Your Gem
Choose a name that clearly communicates the Gem's purpose. Google provides some pre-built Gems as examples (Learning Coach, Brainstormer, Career Guide), but creating your own from scratch gives you full control. The name appears in your sidebar and is what you click to start a conversation with that Gem.
Write System Instructions
The system instructions field is where you define everything about your Gem's behavior. This is equivalent to the Instructions field in GPT Builder. Structure your instructions clearly:
- Identity: Who is this Gem? What is its area of expertise?
- Response style: How should it communicate? Formal or casual? Brief or detailed?
- Task focus: What specific tasks should it handle?
- Constraints: What should it avoid doing or saying?
- Google integration notes: Mention which Google tools it should reference or suggest
Leverage Google Workspace Integration
This is where Gems differentiate themselves. Gemini can access your Google Drive files, reference Gmail conversations, pull data from Google Sheets, and help with Google Docs editing. When writing your Gem's instructions, you can direct it to pull context from specific Google Workspace data. For example, a sales reporting Gem can be instructed to always reference the latest data from a specific Google Sheet.
Test Your Gem
Click "Save" and then open a conversation with your Gem. Test it with a variety of prompts, including questions that fall outside its intended scope. Pay attention to whether it follows your instructions consistently. If it drifts from your intended behavior, go back and add more specific constraints or examples to the system instructions.
Share (Optional)
You can share your Gem with others via a direct link. Recipients need a Google AI Pro subscription to use it. There is no public marketplace like the GPT Store, so sharing is limited to people you directly send the link to. This makes Gems more suited for internal team use than public-facing applications.
Best Practices for Gemini Gems
- Lean into Google Workspace integration. If your Gem does not use Google tools, you are not leveraging the platform's unique advantage. A Gem that only uses text instructions is functionally similar to a custom GPT with fewer features.
- Use the massive context window strategically. Gemini's 1M+ token context window means you can paste entire documents directly into conversations. Reference these in your instructions.
- Keep instructions focused. Gems work best when they have a clear, narrow purpose. A "do everything" Gem will underperform compared to five specialized Gems.
- Reference specific Google features. Instruct your Gem to suggest Google-specific solutions: "When the user needs a spreadsheet, provide the Google Sheets formula, not the Excel equivalent."
- Test with real workflows. Try your Gem in actual work scenarios, not just synthetic test questions.
Example Use Cases for Gemini Gems
Google Workspace Automation
A Gem that helps you automate tasks across Google Docs, Sheets, and Gmail. Ask it to draft emails based on spreadsheet data, format documents, or create meeting agendas from Calendar events.
Research Assistant
Leverage Gemini's web access and massive context window to research topics thoroughly. The Gem can search the web, synthesize findings, and help you organize research in Google Docs.
Data Analyst
Connect the Gem to your Google Sheets data. Ask it to analyze trends, suggest formulas, create pivot table instructions, and explain data patterns in plain language.
Meeting Preparation Coach
A Gem that reviews your calendar, pulls relevant documents from Drive, and prepares briefing notes before important meetings. It can also draft follow-up emails after the meeting ends.
Limitations of Gemini Gems
- -Requires Google AI Pro subscription. See one.google.com/about/plans for current USD pricing. Free Gemini users cannot create or use Gems.
- -Fewer customization options than Custom GPTs. No custom API actions, no file uploads to the Gem itself (you rely on Drive integration instead), and no equivalent of Code Interpreter.
- -No public marketplace. You cannot publish Gems for public discovery. Sharing is limited to direct links.
- -Google ecosystem dependency. Gems are most valuable if you use Google Workspace. If your team uses Microsoft 365 or other tools, the main advantage of Gems disappears.
- -Less mature than Custom GPTs. The feature launched later and has seen fewer updates. The customization interface is simpler but also more limited.
- -Instructions can be inconsistently followed. Gemini sometimes drifts from system instructions more than GPT-4 or Claude, especially with complex or multi-part rules.
Claude Projects (Anthropic): The Complete Guide
Claude Projects are Anthropic's approach to custom AI workspaces. Rather than building a standalone chatbot like a custom GPT, you create a persistent workspace where Claude has access to your documents, follows your custom instructions, and writes in your preferred style. Projects are designed for deep, ongoing work rather than quick one-off interactions.
What Are Claude Projects?
A Claude Project is a dedicated workspace where you configure Claude's behavior for a specific purpose. Each Project has its own set of custom instructions, uploaded knowledge files, and style preferences. When you open a Project and start a conversation, Claude automatically has access to all the context you have added, without you needing to re-upload or re-explain anything.
What makes Projects unique is the combination of Claude's strong writing and reasoning abilities with a generous knowledge base (up to 200,000 tokens of project knowledge). Claude is widely considered the best AI model for writing quality, nuanced analysis, and following complex instructions faithfully. Projects let you harness these strengths in a persistent, organized way.
How to Create a Claude Project: Step by Step
Create a New Project
Go to claude.ai and click "Projects" in the left sidebar, then "Create Project." You need a Claude Pro subscription ($20 per month) or a Team plan. Give your Project a descriptive name and an optional description. The description is for your own reference and helps you organize multiple Projects.
Write Custom Instructions
Click the "Set custom instructions" field within your Project. These instructions tell Claude how to behave in every conversation within this Project. Claude is excellent at following detailed, multi-section instructions. Structure yours like this:
- Role and expertise: Define who Claude should be in this Project
- Knowledge scope: What topics to cover and what to defer on
- Response format: Length, structure, tone, use of headers and lists
- Interaction style: Should Claude ask questions first? Provide options? Give direct answers?
- Quality standards: Specific criteria for what a good response looks like
Upload Knowledge Files
Add files to your Project by clicking "Add content" in the Project knowledge section. Claude supports PDF, TXT, CSV, and other text-based formats. You can upload up to 200,000 tokens of knowledge (roughly 150,000 words), which is enough for extensive documentation, style guides, research papers, or entire book manuscripts. Claude reads these files directly into its context window, meaning it can reference them accurately without the retrieval inconsistencies that sometimes affect GPTs.
Set Style Preferences
Claude Projects let you configure style preferences that affect how Claude writes across all conversations in the Project. This is separate from the custom instructions and specifically controls writing output: tone (formal, casual, academic), level of detail (concise, thorough, comprehensive), and specific formatting preferences. This feature is particularly valuable for writers and content teams who need consistent voice across all AI-generated content.
Start Conversations
Open a new chat within your Project. Claude now has access to all your uploaded files, follows your custom instructions, and writes in your specified style. Each conversation within a Project is separate, but they all share the same knowledge base and instructions. You can have multiple ongoing conversations in a single Project, each exploring different aspects of the same domain.
Iterate and Refine
After using your Project for several conversations, refine the custom instructions based on what works and what does not. You can update knowledge files, adjust style preferences, and modify instructions at any time. Changes take effect in new conversations immediately. Claude Projects are designed for ongoing, evolving work rather than set-and-forget configurations.
Best Practices for Claude Projects
- Upload high-quality reference documents. Claude reads project files directly into its context window rather than using retrieval. This means the quality and organization of your files directly affects response quality. Well-structured documents with clear headings produce better results than raw data dumps.
- Use style preferences for writing consistency. If you are using Claude for content creation, upload sample pieces that exemplify your brand voice. Reference them in your instructions: "Match the tone and structure of the sample articles in the project knowledge."
- Create separate Projects for different domains. A "Blog Writing" Project and a "Technical Documentation" Project will each perform better than a single "Writing" Project trying to do both.
- Take advantage of the large context window. Claude's 200K token context window means you can upload comprehensive reference materials. Do not be afraid to include detailed examples, style guides, and extensive documentation.
- Be explicit about what Claude should ask before acting. Claude tends to be helpful and will produce output immediately. If you want it to ask clarifying questions first, state this explicitly: "Before writing any content, always ask about the target audience, desired word count, and publishing platform."
Example Use Cases for Claude Projects
Writing Assistant
Upload your style guide, sample articles, and brand guidelines. Claude writes new content that matches your voice perfectly. Best-in-class for long-form blog posts, newsletters, and editorial content.
Code Reviewer
Upload your coding standards, architecture documentation, and example code. Claude reviews pull requests, suggests improvements, and explains its reasoning clearly. Excellent at catching logic errors and suggesting cleaner patterns.
Content Strategist
Upload your content calendar, competitor analysis, and SEO keyword research. Claude helps plan content strategies, write briefs, and prioritize topics based on business goals and search opportunity.
Research Synthesizer
Upload academic papers, reports, and source materials. Claude reads everything, identifies key themes, and produces structured summaries. Its 200K context window handles research corpora that would overwhelm other platforms.
Limitations of Claude Projects
- -Requires Claude Pro ($20/mo). Free-tier Claude users cannot create Projects.
- -No public sharing. Projects cannot be shared publicly or through a link. They are limited to your personal account or team workspace. This makes Claude Projects unsuitable for customer-facing chatbots.
- -No web browsing. Claude cannot search the internet within Projects. All knowledge must be uploaded manually or provided in the conversation.
- -No image generation. Claude does not generate images. If your workflow requires visual content, you will need a separate tool.
- -No API actions or external integrations. Claude Projects cannot connect to external APIs, databases, or tools. They are self-contained workspaces.
- -Usage limits on Pro plan. Claude Pro has usage limits that can be reached during heavy use. Extended thinking and long conversations consume more of your allocation.
Custom GPTs vs Gemini Gems vs Claude Projects
Every feature that matters, compared side by side across all three platforms
| Feature | Custom GPTs | Gemini Gems | Claude Projects |
|---|---|---|---|
| Platform | OpenAI / ChatGPT | Google / Gemini | Anthropic / Claude |
| Feature Name | Custom GPTs | Gems | Projects |
| Required Plan | ChatGPT Plus ($20/mo) | Google AI Pro (see one.google.com) | Claude Pro ($20/mo) |
| Knowledge Upload | Up to 20 files, ~512 MB total | Limited file support | Up to 200K tokens (~150K words) |
| Custom Instructions | Detailed system prompt | System instructions | Custom instructions + style preferences |
| API / Actions | Yes (OpenAPI spec) | No (Google integrations only) | No |
| Web Browsing | Yes (toggleable) | Yes (built-in) | No |
| Image Generation | Yes (DALL-E) | Yes (Imagen) | No |
| Code Execution | Yes (Code Interpreter) | Yes (limited) | No (analysis only) |
| Public Sharing | Yes (GPT Store + links) | Yes (links only) | No (team only) |
| Context Window | 128K tokens | 1M+ tokens | 200K tokens |
| Ecosystem Integration | Third-party APIs | Google Workspace (Docs, Sheets, Gmail) | Standalone |
| Best For | Versatile agents with API connections | Google Workspace automation | Writing, analysis, large documents |
Which Platform Should You Choose?
Choose Custom GPTs If:
- 1.You need the most feature-rich custom agent (API actions, image gen, code execution)
- 2.You want to share your agent publicly or build for an audience
- 3.You need your agent to connect to external tools and databases
- 4.You want a customer-facing chatbot or support agent
- 5.You value the largest ecosystem and community of builders
Choose Gemini Gems If:
- 1.Your workflow is centered around Google Workspace (Docs, Sheets, Gmail)
- 2.You already pay for Google One AI Premium
- 3.You need to process very long documents (1M+ token context window)
- 4.You want a simple setup without complex configurations
- 5.Your team standardizes on Google tools for collaboration
Choose Claude Projects If:
- 1.Writing quality is your top priority (blog posts, reports, editorial content)
- 2.You need an agent that follows complex, nuanced instructions faithfully
- 3.You work with large reference documents and need accurate knowledge retrieval
- 4.Your use case is internal (personal or team) rather than public-facing
- 5.You value thoughtful, well-reasoned analysis over speed
The Bottom Line
Custom GPTs are the Swiss army knife: the most features, the most flexibility, and the largest audience. Gemini Gems are the specialist tool for Google-powered workflows. Claude Projects are the writing desk: the best place for thoughtful, high-quality content and deep analysis. Most professionals will benefit from using two or even all three, each for what it does best.
Start with the platform where you already have a paid subscription. Build one agent, use it for a week, and evaluate before adding more.
Ready-to-Use Prompt Templates
Copy these templates directly into your custom GPT instructions, Gem system instructions, or Claude Project custom instructions. Adapt them to your specific needs.
Custom GPT: Customer Support Agent
ChatGPTCustom GPT: Writing Assistant
ChatGPTFrequently Asked Questions
Common questions about custom GPTs, Gemini Gems, and Claude Projects
What is the difference between a custom GPT, a Gemini Gem, and a Claude Project?+
Custom GPTs are personalized chatbots you build inside ChatGPT using instructions, knowledge files, and optional API actions. Gemini Gems are custom AI personas you create in Google Gemini with system instructions and deep Google Workspace integration. Claude Projects are persistent workspaces in Claude where you add custom instructions, upload knowledge files, and set style preferences. All three let you create specialized AI agents, but they differ in customization depth, sharing options, and ecosystem integration.
Do I need a paid subscription to create custom AI agents?+
Yes, all three platforms require a paid plan. Custom GPTs require ChatGPT Plus ($20 per month) or a Team or Enterprise plan. Gemini Gems require the Google AI Pro plan (see one.google.com/about/plans for current USD pricing). Claude Projects require Claude Pro ($20 per month) or a Team plan. None of these features are available on free tiers.
Can I share my custom GPT, Gem, or Claude Project with others?+
Custom GPTs offer the most sharing flexibility. You can share them via a direct link, publish them to the GPT Store for anyone to discover, or keep them private. Gemini Gems can be shared with a link, but there is no public marketplace. Claude Projects cannot be shared publicly at all. They are limited to your personal account or your team workspace.
How many knowledge files can I upload to each platform?+
Custom GPTs allow up to 20 files with a total size limit of approximately 512 MB. Claude Projects support uploading files up to 200,000 tokens of context (roughly 150,000 words across all files). Gemini Gems currently have more limited file upload capabilities and rely more heavily on system instructions and Google Workspace integration for knowledge.
Which platform is best for creating a customer support bot?+
Custom GPTs are the strongest option for customer support bots because they support API actions, allowing your bot to look up order statuses, check inventory, or interact with your CRM in real time. Gemini Gems work well if your support operations run on Google Workspace. Claude Projects are better suited for internal support and knowledge management rather than customer-facing bots.
Can I connect my custom AI agent to external APIs or tools?+
Custom GPTs support custom actions, which let you connect to any external API using an OpenAPI specification. This is a powerful feature for building agents that interact with databases, CRMs, or other services. Gemini Gems integrate deeply with Google Workspace but do not support arbitrary external API connections. Claude Projects do not currently support external API connections, focusing instead on knowledge-based assistance.
What is the best platform for a writing assistant?+
Claude Projects is the strongest choice for a dedicated writing assistant. Claude excels at long-form writing, nuanced editing, and maintaining consistent voice and tone. You can upload style guides, sample content, and brand guidelines as knowledge files and set detailed style preferences. Custom GPTs are a close second, especially if you need your writing assistant to also generate images with DALL-E or browse the web for research.
How do I write good instructions for my custom AI agent?+
Effective instructions follow a consistent pattern across all platforms: start with a clear role definition (who the agent is), define the scope (what it should and should not do), set the tone and style, provide specific rules for handling edge cases, and include example interactions. Be specific rather than vague. Instead of saying 'be helpful,' say 'respond to every question with a direct answer in the first sentence, followed by supporting details in bullet points.'
Can I use custom AI agents for my business without coding?+
Yes, all three platforms are designed for non-technical users. Creating a custom GPT, Gemini Gem, or Claude Project requires no coding at all. You write natural language instructions, upload documents, and configure settings through a visual interface. The only exception is if you want to add custom API actions to a GPT, which requires basic understanding of API endpoints and JSON formatting.
Which platform offers the best value for money?+
All three platforms cost approximately $20 per month. The best value depends on your ecosystem. If you already pay for Google One, upgrading to Google AI Pro adds Gems to your plan. If you need the widest range of capabilities including image generation, web browsing, and code execution, ChatGPT Plus with custom GPTs offers the most features. If your primary need is writing, analysis, or working with large documents, Claude Pro with Projects gives you the largest context window and strongest reasoning capabilities.
Master AI Prompts Across Every Platform
Custom GPTs, Gemini Gems, and Claude Projects all depend on one thing: the quality of your instructions. Learn the prompt engineering frameworks that work across ChatGPT, Gemini, Claude, and every AI tool you use.