I wouldn’t say that my prompts are more proper than anyone else’s, but GPT-5.4 adheres to them really well and does a good job of asking clarifying questions when needed. Spending some time up front to make a very detailed list of requirements here really helped. I think Claude Code still creates better UIs when given a broad task, but Codex makes totally usable interfaces when it’s fed designs and specific instructions. Your mileage may vary; I think different people’s development styles may mesh better with different models and harnesses. I don’t think that Claude Code is a clear winner in April 2026 compared to how it outperformed other options a few months ago.

I was largely inspired by beads (https://github.com/gastownhall/beads) but ended up vibe coding something slightly more tailored for the way I like to work. It’s basically a JIRA style task with a few fields (project, title, status, description, priority, parent task, and an array of notes) that I store in a Postgres database hosted in Neon. I have a lightweight API and human readable UI hosted in fly.io. The CLI is a Python API client that gets executed with uvx and can retrieve a list of tasks in JSONL format, get task details for a specific task ID, create a new task or update an existing task. I defined a skill for Claude Code and Codex to use the CLI and was off to the races- both harnesses are very good at following instructions to use the task tracker to manage work items.

This approach works well for me because I can easily view the task list and notes from any device and coding agents can coordinate across repos or development environments. In hindsight it might have been easier to wrap something like Todoist with a CLI so I didn’t have to host any of my own infra. I do think writing your own task manager with a coding agents is almost a standard right of passage in 2026- if you’re going to do that you might as well make one that you can use to coordinate work with LLM coding tools.