AI Model Cheat Sheet: Claude, ChatGPT, and Gemini

Ian Lurie, May 13, 2026

Agents, Gems, GPTs… I sure wish all these platforms used the same vocabulary. For now, though, I’ve tried to put together a page that tells you which feature on which platform does what. When the underlying feature is the same across platforms, the description is the same too — so you can spot at a glance where the platforms genuinely match up.

Core concepts

Concept Claude ChatGPT Gemini
Reusable assistant
Something repeatable. Think instructions, context, and rules that you can invoke in multiple sessions to produce consistent results.
Skill A packaged assistant with custom instructions and uploaded resources (optional), reusable across chats. Custom GPT A packaged assistant with custom instructions and uploaded resources (optional), reusable across chats. Gem A packaged assistant with custom instructions and uploaded resources (optional), reusable across chats.
Agents
Agents perform multi-step tasks. This is what most of us think of when we think “AI.” You give it an instruction, and the agent goes and works its magic, returning a result.
Workflow / Cowork Cowork (a desktop tool for non-developers) or Claude Code (for developers and the nerds among us). Claude Desktop does similar stuff. Drives tasks across files, connectors, and code. Agent Mode Fires on a task-by-task basis from the composer. Plans and executes using a virtual browser, code execution, and connectors. No single named feature The closest thing to “agents” is a combination of Deep Research and Canvas. You tie them together.
Persistent workspace
A set of instructions, files, and a chat history the model can use to carry out a specific type of task without re-explanation.
Project A workspace with custom instructions, uploaded files, and its own separate chat history. Context stays within the project. Project A workspace with custom instructions, uploaded files, and its own separate chat history. Context stays within the project. Gem + Drive, or NotebookLM Gemini doesn’t have “Projects” per se. They have a combination of Gems, Drive access, and NotebookLM. Don’t let the multiple tools scare you. It’s very, very powerful.
Custom instructions
The “system prompt” we’ve all heard so much about. They apply to every chat, no matter what.
User preferences / Styles Global instructions about you and how to respond. Applies to every chat unless overridden inside a Skill or Project. Custom instructions Global instructions about you and how to respond. Applies to every chat unless overridden inside a Custom GPT or Project. Saved info Global instructions about you and how to respond. Applies to every chat unless overridden inside a Gem.
App and service connectors
How you connect to other services like Google Drive. Doesn’t include MCP servers or custom API connections.
Connectors First-party connections to external apps. Connectors / Apps First-party connections to external apps. Workspace integration / Extensions It’s Google, so native integration with Gmail, Docs, Sheets, Drive, Calendar, and Meet built in. Otherwise, first-party connectors to external apps.
Command-line agents
Direct, non-GUI access to the AI. It’s a great nerd flex, and it’s the most powerful way to access the model.
Claude Code Command-line tool that drives the model from the terminal. Also has IDE integration with VS Code and other tools. Codex CLI Command-line tool that drives the model from the terminal. Gemini Code Assist / Gemini CLI Command-line tool that drives the model from the terminal. Also has IDE integration with VS Code and other tools.
Project context file (for command-line agents)
A markdown file placed in your project root or in the directory tree. The CLI agent reads it before performing a task. This is the CLI agent’s superpower. Learn to use it!
CLAUDE.md A markdown file the agent loads before it executes any task within that directory tree. Note that Claude will accept AGENTS.md as well. AGENTS.md A markdown file the agent loads before it executes any task within that directory tree. GEMINI.md A markdown file the agent loads before it executes any task within that directory tree.

 

This is how I teach the shit out of everything. Got questions? Reach out to me! Seriously, I love to help out:

Or, leave a comment below: