Building Mama Flow with Lovable, ChatGPT, and Codex: speed, cost, and quality
Mama Flow is a B2C product supporting expecting mothers throughout pregnancy with personalized weekly letters. It is also the project where I really learned to combine several AI tools in the same software delivery chain.
Lovable quickly brought me back into a product-building loop
What attracted me to Lovable was not code generation alone. It was how quickly an idea for a user journey could become something visible, testable, and editable.
For Mama Flow, that short loop has been particularly useful. The product combines acquisition, onboarding, personalization based on due date, a user space, subscriptions, a back office, and day-to-day operations. Materializing product decisions quickly makes it possible to challenge them earlier.
But that speed also creates a temptation: keep asking the same tool to do everything.
One tool for everything is rarely the most efficient option
Over time, I started separating three roles.
Lovable remains very effective for visual changes and product iterations that benefit from immediate feedback.
I use ChatGPT more to step back: challenge a flow, work on content, analyze architecture, prepare a specification, or think through a testing strategy.
Codex takes over when the task needs direct repository work: cross-cutting changes, technical quality, checks, consistency across files, or work that benefits from a more deterministic coding environment.
This is not a universal rule. It is simply a way to avoid paying several times for the same reasoning and to choose the tool based on the actual job.
Cost optimization is mostly about distributing context better
Generative tools consume more when iterations remain broad and ambiguous.
I gradually moved long-form reasoning, analysis, and writing outside Lovable, then send it more focused requests when the goal is a concrete visual or functional change.
When a change affects several files, security rules, or technical invariants, I prefer to prepare a specification and let Codex work directly in the repository.
That separation improves cost, but more importantly it improves feedback quality. Each tool receives context adapted to the work it is good at.
AI does not remove the requirements of a production product
Mama Flow is not a prototype. The service has real users, subscriptions, payments, personal data, notifications, analytics, and a back office.
At that point, “it works in the preview” is no longer enough.
The usual software product concerns still apply: security, identity management, permissions, migrations, observability, analytics, cost control, non-regression, and the ability to understand what happened when something goes wrong.
AI accelerates many tasks. It removes none of those responsibilities.
Mama Flow moved me from prompt user to orchestrator
At first, it is easy to judge an AI tool by the quality of its immediate answer.
With a real product, the question changes: how do several tools, contexts, and levels of autonomy work together to produce a reliable outcome?
I therefore spend less time looking for “the best prompt” and more time working on the system around it: who prepares the decision, who changes the code, who checks it, which rules persist in the repository, and which controls prevent regression.
That is exactly the kind of capability I want to keep developing as a CTO / CTPO.
The product remains the judge
Additional speed only matters if it helps us learn faster from users without degrading their experience.
Mama Flow supports expecting mothers throughout pregnancy with 34 weeks of personalized letters adapted to their stage, helping them navigate pregnancy more calmly and prepare for their baby’s arrival.
That product promise matters more than the tools used to build it.