I built a Chrome extension in 40 minutes, from idea to working extension. No modifications were necessary after testing.
The extension exports a complete ChatGPT conversation to a clean markdown file with numbered turns and speaker labels. You open a chat, click the extension, and it captures everything. The scope was simple, the constraints were clear, and there was a single deliverable.
Three distinct AI systems from two companies handled three different roles. I collaborated with ChatGPT to develop the intent. I provided the concept, constraints, and scope boundaries. ChatGPT asked clarifying questions and wrote the intent document. After nine exchanges over about ten minutes, the spec was finalized. I was the Architect, and ChatGPT was the pen.
Codex acted as the Manager. It read the intent and generated the run bundle: a task list, a developer prompt, and operating instructions. Seven tasks, clearly scoped, with acceptance criteria directly based on the intent. About ten minutes.
Claude Code acted as the developer. It read the run bundle and executed it. Seven tasks completed with zero failures and no errors. In about ten minutes, it produced a working Chrome extension. The last ten minutes were spent testing. The extension functioned as specified, with no changes needed.
That breakdown is intentional, not accidental. Spend ten minutes defining the intent, ten minutes planning the execution, ten minutes building, and ten minutes verifying. The work is done upfront. Once the decisions are made and documented, execution becomes simple.
The choice of tools didn't matter because Trail remained consistent regardless of which AI filled which role. Three systems, two companies, different models. The framework succeeded because roles and handoffs were clear, and the intent was unambiguous going in, so there was no need for guesses later. The choice was intentional, not random. Each model was chosen for its strengths: Codex generated clearer planning artifacts and more organized developer instructions, while Claude Code excelled at interpreting intent and producing a usable, well-constructed result. The pattern is simple: use the model suited for the role, not your personal preference.
There is one constraint that outweighs tool choice. The Developer must never run in the same conversation where the intent was created. Doing so allows the model to inherit context that isn't present in the artifacts, violating Trail’s most important rule: the Developer’s context must match the file context. If the Developer knows something not documented, the system can no longer be trusted. Use separate windows, separate contexts—no exceptions.
Web: trail.venturanomadica.com
GitHub: github.com/Ventura-Nomadica/trail-framework
Top comments (0)