I would like to share my experience as a sole founder building software products using ChatGPT, Codex and Visual Studio Code, and propose an improvement that I believe could make a significant difference to independent developers and founders.
When OpenAI first introduced Codex into my development workflow, one of the things I particularly valued was that, after receiving a usage-limit warning, Codex could continue working towards completing the task it had already started.
More recently, my experience has been different. Codex can stop in the middle of an engineering assignment when its usage allowance runs out.
I understand that computational resources are expensive and that usage limits are necessary. I am not asking OpenAI to provide unlimited Codex usage.
What I am asking is for OpenAI to protect the completion and continuity of work that has already been started.
How I use ChatGPT and Codex as a sole founder
I am building products without the luxury of having separate, fully staffed product, design, engineering, research and quality-assurance departments.
Instead, I use AI to help me perform those responsibilities.
ChatGPT acts as my product strategist, product designer, researcher, technical adviser and independent reviewer. Together, we deliberate, research, design capabilities, examine user journeys, identify blind spots and determine what should be built.
Codex acts as my engineer. It implements the agreed designs, examines the existing source code, runs tests, identifies and fixes bugs, checks for regressions, captures screenshots and produces detailed reports and downloadable evidence packages on my local computer.
I then act as the Founder and final reviewer.
I manually test the functionality, examine the interface, compare the implementation against the approved visual references and experience the product from the perspective of the people who will eventually use it.
I decide whether the work should be accepted, refined or rejected.
I also work directly in the Visual Studio Code terminal with instructions from ChatGPT, particularly when Codex is unavailable or when a smaller, controlled correction is required.
These are not isolated activities. They are interconnected responsibilities within one development operation.
The evidence produced by one department enables the next department to do its job.
What happens when Codex stops halfway through its work?
When Codex reaches its usage limit before finishing an assignment, I may not receive the screenshots, implementation reports, test results or downloadable evidence packages that I would ordinarily receive at the end of the task.
Without those screenshots, I cannot properly compare the implemented interface against what was originally designed.
Without the completed report, I cannot establish whether everything was implemented, whether something was missed, whether a regression was introduced or whether the necessary tests were completed.
Without that evidence, ChatGPT cannot necessarily provide a reliable independent assessment or determine the correct next action.
This becomes particularly important when I want to continue working directly in the Visual Studio Code terminal while waiting for my Codex allowance to reset.
ChatGPT may not have access to the latest uncommitted source changes, the isolated development environment, the screenshots or the exact implementation state at which Codex stopped.
Consequently, instead of moving forward, we may have to spend considerable time investigating the source code, examining terminal outputs, reconstructing the assignment and determining what Codex did before it stopped.
Sometimes, this investigation consumes additional AI usage simply to recover work that was already underway.
As someone who is responsible for the entire product, I find this particularly frustrating because the responsibility for restoring continuity falls back on me.
User experience is more than passing engineering tests
For me, building a successful software product is not simply about writing code or getting automated tests to pass.
It is about how the product feels, how people interact with it, whether it solves their problems, whether its functionality is intuitive and whether the experience encourages them to return.
A feature can pass its engineering tests and still fail to deliver the intended user experience.
This is why screenshots, visual comparisons, browser testing and detailed implementation evidence are essential to my workflow.
I need to be able to compare what was designed with what was actually built. I need to examine whether previously approved functionality has been preserved and whether the new implementation has introduced unexpected problems elsewhere.
When Codex stops before producing this evidence, the quality of the product can suffer.
The next assistant may understand the technical problem but miss the original design intention. Previously approved refinements may be overlooked. A correction may unintentionally introduce a regression.
Over time, the product can drift away from the carefully considered experience that the Founder originally intended.
This is not because the Founder has changed the vision. It can happen because the continuity between design, implementation, review and testing has been interrupted.
Why the previous behaviour was so valuable
When Codex could finish an existing assignment after a usage-limit warning, I could receive the completed work, reports, screenshots, test evidence and recommendations.
Even though I might have to wait several hours before starting another Codex task, I could still use that time productively.
I could continue working with ChatGPT, independently review the implementation, manually test the product, research improvements, refine the user experience and prepare the next development sprint.
I could also carry out smaller corrections directly in my Visual Studio Code terminal, using the completed engineering report as a reliable reference.
The work did not have to stop simply because one part of my AI team had temporarily reached its usage allowance.
That was an exceptionally useful way of working, particularly for someone building businesses independently.
The financial reality of independent founders
Not everyone developing ambitious software products has investors, substantial funding, an established engineering team or a business that is already generating revenue.
Some of us are building our products from scratch, investing our own money, time, energy and creativity long before we receive any financial return.
Purchasing additional Codex credits or upgrading to a more expensive subscription may provide a larger allowance, but it does not necessarily guarantee that a complex engineering assignment will reach completion before another limit is encountered.
The underlying problem therefore remains.
I believe one of the greatest opportunities AI offers is the ability for ordinary people with ambitious ideas to build businesses that previously would have required substantial capital and large professional teams.
That opportunity becomes harder to sustain when independent founders must repeatedly increase their spending or devote significant time to recovering interrupted work.
AI should help reduce the financial and operational barriers to innovation, not unintentionally introduce new ones.
My proposal: Codex Task Completion Protection
I would encourage OpenAI to consider the following improvements.
1. Allow an already-running task to reach completion
When a user reaches their usage limit during an active engineering assignment, provide a reasonable, controlled completion allowance.
This would apply to the existing authorised assignment only. New assignments could remain unavailable until the normal usage allowance resets.
2. Treat verification and evidence as part of task completion
Completion should not simply mean that Codex has finished modifying the source code.
It should also include the testing, regression checks, screenshots, visual comparisons, implementation reports, downloadable evidence packages and recommendations required by the original assignment.
These deliverables are essential to the people and AI assistants responsible for the next stage of work.
3. Provide an intelligent warning before starting substantial work
If Codex estimates that the remaining usage allowance may be insufficient for a proposed assignment, it should inform the user before beginning.
This could help developers make informed decisions about the scope of work and reduce the likelihood of interrupting critical implementation.
4. Guarantee a meaningful engineering handoff when completion is impossible
I recognise that some assignments may be unusually large, encounter unexpected technical problems or exceed a reasonable completion allowance.
In those circumstances, Codex should automatically produce an editable and downloadable recovery package containing:
- The original assignment and approved requirements.
- The files examined, created or modified.
- What was implemented and what remains outstanding.
- Tests passed, failed and not yet performed.
- Available screenshots and browser-testing evidence.
- Known defects, regressions and unresolved problems.
- The exact repository, branch and implementation checkpoint.
- Relevant terminal commands and their results.
- Safe recovery or rollback options.
- The exact recommended next action.
This should be saved locally where appropriate, so the Founder can provide it to ChatGPT, another authorised assistant or a human engineer.
The user should not have to reconstruct the assignment or search through extensive terminal logs merely to determine where work stopped.
5. Protect users from unexpected expenditure
This mechanism should not automatically purchase additional credits or incur charges without the user’s consent.
OpenAI could introduce reasonable fair-use protections or make this available initially to paying subscribers and users engaged in sustained software development.
The intention is not to remove usage limits but to make them less disruptive to professional work.
Clarification of the current behaviour
OpenAI’s Help Center describes circumstances in which an active Codex turn may continue after a usage limit is reached, subject to fair-use limits.
However, my recent experience has been that active engineering assignments can stop before completion.
I would appreciate clarification about whether the continuation behaviour has changed, whether different restrictions now apply or whether there is a technical issue affecting this experience.
There is also an important distinction between completing an AI model’s turn and completing the engineering assignment that the user actually authorised.
For developers and founders, it is the completion of the meaningful assignment that matters.
Why this matters beyond usage limits
I have previously raised a related feature request concerning automatic conversation handoffs in ChatGPT when long-running conversations reach their maximum length.
Both issues expose a wider challenge in AI-assisted professional work: continuity.
When ChatGPT loses the history of why a product was designed in a particular way, and Codex stops before documenting how that design was implemented, the Founder becomes responsible for reconnecting everything.
The very technology we adopted to accelerate our work can unintentionally create another administrative burden.
I want to spend my time creating, innovating, solving problems, improving the experience of my future customers and bringing my products to market.
I do not want to spend increasing amounts of time reconstructing conversations, investigating unfinished AI work, repeating established decisions or correcting regressions caused by incomplete handoffs.
These interruptions affect more than productivity. They affect creative momentum, financial planning, confidence, motivation and, ultimately, the ability of independent founders to bring their ideas to life.
I believe OpenAI has an opportunity to make ChatGPT and Codex not only powerful individual tools, but dependable professional working partners that support continuity across the entire development process.
AI should not be measured only by how much work it can start, but by how reliably it helps people finish what they have started.
I would welcome OpenAI’s product and engineering teams considering this proposal, as well as hearing from other developers and independent founders who may be experiencing similar challenges.
Thank you for reading.