How tool-use is democratizing AI

22 May, 2025

You don't have to work at Anthropic or OpenAI to work on world-leading AI. While they focus on model development, that's only a small part of the picture. How you use the model is much more important.

My early experiences with ChatGPT-like solutions showed promise - helping me code 1-2x faster. But I wasted so much time copying and pasting code into chat windows. For complex tasks spanning multiple files or requiring context from your codebase, these tools often fell short. The LLMs seemed too disconnected from real-world development. Everything changed when I started experimenting with Claude Code (CC). While not the first tool-based LLM solution, it demonstrated something revolutionary - the same model that struggled before could now solve complex problems in one shot when equipped with the right tools. The magic wasn't in some special sauce - it was in giving the AI access to your actual development environment through tools. Simple capabilities like viewing files and folders, running bash commands, searching codebases, and making API requests transformed how effectively the AI could assist.

This insight led me to create Merlin, our CLI tool that helps engineers with software tasks. By starting with a solid foundation of tools and the right prompting, we achieved remarkable results. We expanded that functionality way beyond just a CLI tool to automatically convert standups into actionable items, transform demos into operational documentation, generate and maintain onboarding documents that stay current, and automate release notes by analyzing diffs and Jira tickets.

The results were dramatic - tasks that used to take 3 hours could be completed in 20 minutes. And we're not talking about just generating boilerplate code. Merlin produces senior-level code that follows team code conventions perfectly, creates proper abstractions, isolates logic appropriately, and actually runs without errors.

The core concept is remarkably simple. Instead of asking an LLM to write code directly, you give it access to tools (functions it can call), let it choose which tools to use based on your request, feed results back for the next decision, and repeat until the task is complete. For example, when writing a new feature, Merlin might search the codebase for similar patterns, create new files following team conventions, write the backend logic, generate API endpoints, update frontend code to consume the API, and add tests - all while maintaining perfect context of what it's building across files and systems.

Using this approach, we've seen consistent 10x productivity improvements across our engineering team. Not just in terms of speed, but in the quality and maintainability of code produced. The power of tool-use democratizes AI development. You don't need custom models or massive compute resources. With the right tools and prompts, existing models can achieve remarkable results that were previously thought impossible.