An engineer at the AI-powered app development platform Lovable said in a LinkedIn post that a series of changes to the company’s system prompt helped cut annual large language model (LLM) costs by nearly $20 million, while also improving performance.
Benjamin Verbeek, a member of technical staff at Lovable, explaining that he spent the holiday period reviewing and improving the platform’s system prompt. According to Verbeek, the changes made Lovable about 4% faster.
“The crazy part is that this also ended up decreasing our LLM costs by $20M per year—with help from my mother,” Verbeek wrote.
He explained that when his mother—who is a historian by training—asked him to explain his work in the LLM space. He showed her an LLM trace, which is a detailed record of how an AI model processes instructions and generates responses step by step.
Looking at the trace, Verbeek said his mother asked why certain instructions were repeated multiple times across different parts of the prompt.
“What we realised is that our system prompt is constructed dynamically from lots of different files,” Verbeek added. “As we’ve been optimising each part, no one had looked at the coherence for a while. Together we found duplication, inconsistencies, and overly verbose formulations.”
He explained that over time, engineers had kept adding new instructions to emphasise specific behaviours, without removing or consolidating older ones.
This led to unnecessary repetition and diluted the prompt’s overall effectiveness.
Verbeek said the team removed duplicate instructions, tightened the language, and preserved the original intent and balance of constraints.
After manually rewriting the first sections, he used an AI model to refactor the remaining portions in the same style, followed by a detailed line-by-line review to reintroduce a few critical safeguards.
The revised prompt was then A/B tested over the New Year period.
According to Verbeek, the updated system followed instructions more reliably, responded faster, and significantly reduced token usage, leading to substantial cost savings at scale.
Reflecting on the experience, Verbeek highlighted three key takeaways.
Firstly, he said that prompt quality compounds at scale. He also noted that fresh perspectives can outperform simply “prompting harder.” Lastly, he pointed out that fast, safe experimentation is a major advantage in AI development.
However, one user pointed to a recent Google research paper suggesting that prompt repetition can, in some cases, improve the performance of large language models without increasing output length or latency.
Responding to the comment, Verbeek said that repetition can indeed be beneficial in certain contexts, but stressed that Lovable’s advantage lies in its ability to run high-confidence experiments to validate such claims. “In our conditions, this happened to work really well,” he said.
Last December, Lovable raised $330 million in a Series B funding round at a valuation of $6.6 billion.
Lovable said the capital will be used to deepen integrations with enterprise software tools, expand collaboration and governance features for teams, and strengthen infrastructure that supports moving products from prototype to production.
The post Lovable Engineer Saves Company $20 Mn Yearly With Help From His Mother appeared first on Analytics India Magazine.