The Bitter Lesson of LLM Extensions
2 hours ago
- #Customization
- #LLM Extensions
- #AI Agents
- The evolution of LLM extensions from simple text inputs to complex agents with customizable tools.
- ChatGPT Plugins (March 2023) introduced the idea of LLMs using REST APIs, but models like GPT-3.5 struggled with API specs.
- Custom Instructions (July 2023) simplified repetitive context setting by appending user-defined prompts to every chat.
- Custom GPTs (November 2023) bundled personas, files, and actions into shareable links, moving towards single-purpose apps.
- Memory in ChatGPT (February 2024) enabled automatic personalization by remembering user details across conversations.
- Cursor Rules (April 2024) integrated custom instructions directly into code repositories via `.cursorrules` files.
- Model Context Protocol (MCP) (November 2024) provided a robust framework for models to interact with tools and resources.
- Claude Code (February 2025) introduced multiple extension mechanisms like `CLAUDE.md`, MCP, slash commands, and hooks.
- Agent Skills (October 2025) simplified tool integration by using markdown files and scripts, reducing context bloat.
- The future of LLM extensions may focus on natural language and general-purpose tools, moving away from specialized protocols like MCP.