1% AI
In the beginning of 2025, the engineering team at The Langston Company was like many others. We were skeptical of the grandiose claims about AI's coding abilities from major AI companies and LinkedIn influencers. Every day, we saw headlines and blog posts claiming that "coding is dead" and "AI will replace all software engineers by the end of the decade."
We would try using Copilot or ChatGPT to build new features, and it consistently felt like a smoke screen. To us, AI was mostly useful for generating Stack Overflow–style answers tailored to our questions. That's a helpful tool, but not one that felt like it would revolutionize software engineering.
1%
In early 2025, we were writing between 1%–10% of our production code using AI. We primarily used Copilot in our editors (VSCode and WebStorm), along with ChatGPT and Claude Code in the browser. We provided each tool with a few paragraphs of context detailing what our company does, the programming languages we use, example database schemas and migrations, our tech stack, and some coding style and rules.
Copilot understood our repositories best due to its deep integration with our editors. It could analyze each codebase independently and adapt to our style.
Workflow
Our typical workflow started with planning new features at the database layer. Then we moved to API endpoints, and finally to the frontend. We would occasionally use an LLM to help design schemas, but most of that work was still done by hand. In our experience, LLMs were good at writing query languages, so we used them to generate schema and migration files. The results were usually close based on our provided context and a few pasted examples, but still required manual cleanup. We would copy and paste the output into our editor or let Copilot apply changes. It worked, but it wasn't a great experience. It felt like a slightly better version of Stack Overflow.
We would then move on to the API. We scaffolded endpoints by hand and wrote most of the service layer ourselves. Good old-fashioned software engineering. We used LLMs for complex queries, debugging, and writing unit tests. Still, only a few lines per PR were actually generated by AI.
Lastly, we built the frontend to support the new feature. Our stack includes React, Next.js, TypeScript, and ShadCN. This integrates well with Vercel's ecosystem, especially v0. We would prototype more complex UI in v0 by iterating through prompts. The generated code wasn't great and was often difficult to integrate, so we treated it more as a design tool. We'd take the UI/UX direction from the mockups and build proper components ourselves.
Thoughts
This was a standard web development workflow for over a decade: define your schema, build your CRUD endpoints, and create a UI that interacts with those layers. We used AI tools around the edges, but we didn't trust them with anything substantial. We believed the best, and really the only, way to write software was hands-on, using AI only for trivial tasks. It helped automate unit tests, iterate on simple UI, or answer obscure questions about things like CSS properties.
That mental model held up for most of the year. AI was a tool, not a transformational technology. Around October of 2025, we started experimenting and studying how experienced developers were using these tools in practice.
A year later, we've completely changed how we write software. We've gone from 1% to 99%. Agentic AI workflows have changed how we work, how we solve problems, and how we think about software.
