As a developer, I’m always looking for tools that make my life easier. However, with the never-ending stream of updates Codex provides, I hadn’t taken the time to see what this harness could do for me until I watched a demo of Jason Liu’s setup (developer experience engineer at OpenAI). That demo made me realize there were three headaches I could address right away: communication, project management, and industry updates.
When I see a slew of emails, Slack messages, pull requests, and Teams messages, I feel overwhelmed by the avalanche of things I need to review and respond to. At the same time, I want to stay on top of everything, especially tech industry news. But repeatedly hearing about the “next big AI trend” makes keeping up feel impossible. Adopting agentic development has also split my focus more than ever, tempting me to offload too much of my thinking onto the agent. So, to shrink the noise, here are some ways I use Codex to fix these issues.
The Workspace: One Place for Each Area of Work
Jason Liu demonstrated his personal setup using this monorepo template. What immediately caught my attention was his master vault: a single place for all his projects and workspaces, managed by a “Chief of Staff.”
I adopted the same structure for my setup:
– Vault: All projects, workspaces, experiments, and artifacts created by skills
– Projects: My current work projects
– Industry: News sources and daily tech industry updates
– Work: Email and Slack management
At the root of my vault, I always keep a session open for my Chief of Staff. This is where I define the behavior I want all my agents to share. For example, I ask them to return succinct answers that focus on the key points I need. When an agent grills me on what I know, I’ve instructed it to use an unknown matrix derived from Thariq’s post. If I want every session to have access to a specific skill, I ask the Chief of Staff to add it across the vault.
Each project then adds its own boundaries. Its AGENTS.md and README.md files and gives the agents a strict context scope and define how they should behave within that project. Once each area has clear boundaries, I decide which recurring tasks Codex should handle for me.
Loops: Reducing Cognitive Overhead
Loops are another part of the monorepo template. Each loop runs a cron job with instructions an agent follows. I have extended the idea that loops can be used to reduce repetitive activities and extend my ambitions to sustain breadth and depth.
Every day, an agent scans selected blogs, newsletters, and posts and produce a one-page digest of the main topics, including which groups of sources are discussing them. Instead of trying to read every source, I skim the digest and choose which posts are worth exploring in greater depth. This keeps me informed without making industry updates feel overwhelming.
For communication, I have another loop that tracks unread Slack messages and emails every weekday morning. It tells me what changed, who is waiting on me, what needs my input and which unimportant or spam emails it can delete.
NOTE: Jason Liu’s setup also offers a skill called “write-like-me-bootstrap” (see in his repo), which allows an agent to respond for you in your tone of voice. But for now, I prefer to keep that human part intact, at least until the response workload becomes too much to handle.
Enhancing My Understanding
I firmly believe that my understanding needs to grow even if the industry pushes us towards building an agentic software factory. To build a fully automated software factory is like saying we will never deal with bugs again. That utopian fallacy is something I don’t want to assume. Therefore, I am always striving to find new ways to sustain my critical thinking skills, and I’ve found Codex to be a powerful learning tool.
First, there’s Appshot. I still have moments when I want to inspect what’s happening manually. Having a keyboard shortcut that captures my current window lets me keep my fingers on the keyboard without interrupting my flow.
Second, there’s Side Chat. Sometimes an agent reasons through a problem using terms I haven’t heard before. With Side Chat, I can start another parallel conversation to satisfy my curiosity without interrupting the main agent.
Finally, there’s the Voice feature. It has become a fantastic rubber duck. If I need to prepare for a presentation, I can verbally collaborate with the agent to build a solid presentation and simulate potential Q&A sessions.
Conclusion
Honestly, the improvements you can make to your daily workflow are endless, especially since these harnesses seem to have updates practically every day. Ever since I watched Jason Liu’s demo, I’ve tried to keep a more open mind because the goal is to recognize bottlenecks and find opportunities to reduce that friction through better processes.
One final tidbit: Codex has not replaced the terminal for me yet. Even though Codex has improved many of my workflows, the terminal still seems to be the best place for development. If enough people are interested in why, I’ll write a follow-up blog post about it.