If you’ve spent any real time with AI coding assistants, you know the pain. You start a session, build something promising, and then ten prompts later the model has no clue what you were doing five minutes ago. Context leaks. It forgets the architecture. You end up repeating yourself or, worse, pasting the entire codebase back in.
Jason Liu has been dealing with this problem head-on, and his approach with Codex is worth paying attention to. He’s not just hacking together one-off scripts. He’s running long, multi-session projects where the AI needs to remember what it did last week.
The Context Problem Nobody Talks About
Most people treat AI coding like a chat. You ask, it answers, you move on. That works fine for small tasks—write a function, debug an error, explain a regex. But once your project spans multiple files, multiple sessions, and multiple days, the model’s limited context window becomes a bottleneck.
Liu’s insight is straightforward: don’t let the model start from scratch every time. He structures his workflow so that Codex retains the important parts of the project state between sessions. This isn’t about some magic prompt. It’s about being deliberate about what you feed back in.
What Actually Works
He uses a combination of techniques that sound obvious once you hear them but almost nobody does consistently:
First, he maintains a running project summary. Not a changelog, not a README—a living document that describes the current state of the code, the decisions made, and the things that are still broken. Before each new session, he feeds this summary into Codex along with the most relevant files. This gives the model a fighting chance to remember where things stand.
Second, he breaks work into phases that fit within the model’s context window. Instead of asking Codex to build an entire feature in one shot, he decomposes it into steps that each have a clear input and output. Each step builds on the previous one, but he controls the handoff, not the model.
Third, he uses Codex’s own outputs as part of the next prompt. If Codex generates a plan or a set of changes, he saves that output and references it later. The model doesn’t have to remember what it said—he just shows it.
This is higher effort than typing “finish the app” into a chat box. But it’s also the difference between a toy project and something you could actually ship.
The Personal Angle
I’ve tried similar approaches with Claude and GPT-4 for writing and coding, and honestly, it’s tedious. But Liu’s method works because he treats Codex as a collaborator with a terrible memory, not as an omniscient brain. You wouldn’t expect a human coworker to remember every detail of a project from weeks ago without notes. Same deal here.
The downside is obvious: it requires discipline. You have to maintain that summary. You have to be thoughtful about what context you include. You can’t just fire off prompts and hope for the best. But if your project is actually important—a tool you use daily, a client deliverable, a personal project you care about—the overhead pays for itself.
Where This Breaks Down
Let’s be honest: this approach has limits. If your project is massive—hundreds of files, complex dependencies, multiple contributors—no amount of summary trickery will save you. Codex still can’t see your entire codebase at once. You’ll hit context walls no matter what.
Also, Liu’s method assumes you know what’s important. If you’re still figuring out your architecture or your requirements keep changing, maintaining that summary becomes a moving target. I’ve been there. It’s frustrating.
But for the sweet spot of medium-complexity projects—a web app, a data pipeline, a custom tool—this is the best I’ve seen. It’s not a silver bullet. It’s just good engineering practice applied to a new kind of tool.
The Takeaway
Codex can do impressive things in isolation. But if you want it to help you build something that lasts longer than a single session, you have to manage the context yourself. Jason Liu figured out a system that works. Steal it. Adapt it. Don’t expect the model to do the remembering for you.
Comments (0)
Login Log in to comment.
Be the first to comment!