The ability to bring sophisticated tooling to life without needing an engineering team is a significant shift. We are moving from a world where building software required professional expertise to one where software can be built and refined by non-coders. The ability to solve your problems with software is now accessible to everybody.
Natural Language Development
Builders can finally engage with coding tools in natural language via tools like Claude Code, Windsurf, Lovable, Cursor, and others. Building in English can get you to a demo state very rapidly, but it isn’t without challenges.
My experience building a tool to probe frontier LLM models taught me that success requires starting with a detailed product requirements document, being specific about what you want, and iterating through bugs. I didn’t do any of this when I started to build the tool, but I would have had a much easier time if I had. I do now, a lot of the time, and when I don’t, I generally get to a point where I wish I had.
The Economics Have Shifted
A single person can now accomplish what previously required a development team. The cost reduction is transformational for founders without coding expertise, bootstrappers, and people who have no desire to commercialize the final work. Solutions can be built and deployed at the speed of articulation, rather than the speed of hiring and project management.
Iteration is now faster and more powerful than ever. You can explore different approaches, test variations, and refine solutions in real time. When I hit errors that I couldn’t resolve with one model, I would switch between different AI coding tools until I found one that understood and resolved the problem.
Working With Agentic Code Tools
One of the methods I rely on most is expressing the end state as a series of connected steps. Breaking down what you want into clear, sequential steps gives the AI coding tools the context they need to understand not just what you’re building, but how the pieces fit together. This approach works particularly well when building complex functionality that involves multiple interconnected components.
Post 3 of a series documenting AI safety research and the tools built to support it.