Developing Rail Ticketing Features With AI

The project’s initial AI setup consisted primarily of a root CLAUDE.md file. More detailed module-level instructions, reusable skills, and additional context management had not been extensively configured. As a result, generated implementations required considerable back-and-forth before they were ready for a pull request.

Generated changes needed to fit the application’s existing architecture and coding conventions. An initial implementation still required engineering judgment to determine whether its approach and impact were appropriate.

The engineer describes reviewing changed files, checking architectural fit, and directing revisions as part of the AI-assisted workflow. Applied to an established codebase, this keeps acceptance of the implementation with the engineer rather than relying on the agent’s initial output.

AI implementation could raise questions that the codebase alone could not answer. Moving forward without resolving those questions risked producing work that did not match the intended feature behavior.

The engineer’s documented workflow uses planning and follow-up questions before and during implementation, with input from the project team or product manager where necessary. This provides a way to resolve uncertainty before accepting the resulting changes.