Prompt design in large language models (LLMs) is a blend of art and science, but until now many key decisions were made without controlled evidence. A recent study, based on a contamination-free synthetic corpus (the 'Book of Veyra'), examines three critical factors: instruction format (markdown, plain text, prose, or tabular), the number of simultaneous instructions a system prompt can handle before compliance degrades, and the amount of context a model can retain before recall and honesty decline. The results, evaluated across five models, reveal surprising findings with direct implications for enterprises seeking to integrate AI efficiently.
In the first experiment, with 960 calls per model, the perfect response rate collapsed to zero when the number of rules reached 80, regardless of format, placement (system prompt vs. user turn), or model. Format showed no clear advantage: neither markdown nor plain text consistently outperformed; in fact, one 35B model favored plain text. Placement produced effects as large as format at the 160-rule limit, but the direction varied by model. This suggests there is no universal recipe and that companies must test and adapt their prompting strategies based on the specific model they use.
The second experiment, with 5,520 calls per model, evaluated recall accuracy, false-premise sycophancy, and absent-fact fabrication across a context ladder from 2k to 512k tokens. Recall accuracy remained near ceiling up to 64-128k tokens, then degraded sharply and format-dependently: in one model, the accuracy spread reached 48 points at 128k tokens. Surprisingly, no fabrication was detected in any of the 5,760 probes, and sycophancy remained below 8.3%. What did increase dramatically near each model's context ceiling was outright refusal to answer, rising from 0% to 79-90%. This behavior is qualitatively different from fabrication or sycophancy and represents an operational risk for applications requiring high response availability.
For a company developing AI-based solutions, these findings are a call to action. It is not enough to craft a well-written prompt; the system architecture, context management, and model selection must be considered. At Q2BSTUDIO, as a software and technology development company, we understand that integrating artificial intelligence into business processes must be based on empirical data, not assumptions. That is why we offer consulting and custom software development services that enable our clients to design robust prompting systems capable of handling hundreds of instructions and extensive contexts without losing reliability.
The research also underscores the importance of infrastructure. Using cloud services like AWS or Azure facilitates scaling prompt experiments and monitoring model behavior in production environments. Furthermore, integration with Business Intelligence (BI) tools such as Power BI allows visualization of performance metrics like refusal rates or recall accuracy, helping make informed decisions about when to trim context or adjust instruction count.
Another relevant aspect is cybersecurity. When a model systematically refuses to answer near its context limit, it can signal unexpected behavior that could be exploited if not managed properly. Therefore, at Q2BSTUDIO we incorporate cybersecurity and pentesting practices into our AI developments, ensuring that AI agents are robust against manipulation and maintain data integrity.
In summary, prompt design at scale is not a trivial task. The results of this study provide evidence-based guidance that every organization should consider. At Q2BSTUDIO, we combine our expertise in custom applications, cloud computing, BI, and cybersecurity to help companies navigate this complex landscape. Whether implementing AI agents that handle hundreds of simultaneous instructions or deploying systems that manage contexts up to 128k tokens without precision loss, our focus is on measurable and reliable outcomes. For companies looking to harness the potential of LLMs without falling into the pitfalls of amateur design, partnering with a specialized technology provider is the key to success.
