Our Take
This is a use-case study, not a product advance—Codex already existed; the story is how one practitioner organized his workflow.
Why it matters
Practitioners building with LLMs often hit a wall: single-prompt contexts collapse on real projects. Seeing a working example of context preservation across multiple turns answers a real operational question.
Do this week
Engineer: audit your current Codex or GPT-4 workflows this week to identify where context resets are costing you rework, then test a stateful prompt structure before scaling to your team.
How Jason Liu extended Codex beyond single prompts
OpenAI published a case study on Jason Liu's use of Codex for multi-turn, long-running coding projects. Rather than treating each prompt as isolated, Liu designed workflows that preserve context, track project state, and allow work to continue across multiple interactions without losing prior decisions or code structure.
The specific techniques Liu used center on maintaining a running state of the project within the prompt context, structuring queries to reference prior work, and managing complexity by breaking long-running tasks into named, trackable phases. OpenAI positioned this as a pattern for developers who want to move beyond single-turn code generation into sustained, multi-step development cycles.
Context management is the actual constraint in production coding
Most developers encounter Codex or similar LLMs as single-prompt tools: write a request, get back code, paste it in. That works for isolated functions. It breaks down on real projects, where decisions compound, dependencies accumulate, and context drift causes code to diverge from the actual state of the codebase.
Liu's approach flips that: he treats the LLM as a conversational partner on a project, where each turn remembers what came before. This is closer to how developers actually work with other people. The catch is that maintaining state manually requires discipline—you have to encode project context into every prompt, which adds friction unless you build tooling around it.
This matters now because teams are moving past proof-of-concept AI coding tools and into production systems. The question shifts from "can the LLM write code?" to "can the LLM stay coherent across a real sprint?" Liu's pattern suggests the answer is yes, but only if you design the workflow explicitly for it.
Start by auditing where your context breaks today
If you are using Codex or GPT-4 for coding work, identify the moments where you have to repeat context or rephrase decisions because the LLM lost track. Those moments are friction points where a stateful workflow would save time.
The pattern is straightforward: structure your prompts to include a brief project state summary at the top, reference decisions made in prior turns, and number or label phases so the LLM can track where it is in the larger task. Test this on one small multi-turn project before rolling it into your team's standard practice.
This is not about waiting for better tooling. You can do it now with any LLM that supports longer contexts. The benefit is real if your projects have more than two turns of interaction.