A developer tested the impact of seven common prompt additions on the Claude Code agent using the Haiku 4.5 model. The study involved 156 test runs across five Python coding tasks, including refactoring and logic problems. The results showed that prompt engineering, such as assigning a role or offering tips, significantly influenced the agent's search behavior in specific scenarios.
Assigning the role of a 'senior Python developer' caused the agent to narrow its search scope to Python files only, leading it to miss references in README files. Conversely, phrases like 'this is very important' or 'think carefully' encouraged the agent to broaden its search. Offering a 200 USD tip improved performance in some cases but increased token usage and costs by up to 19 percent.
The study concluded that explicit instructions are more effective than role-playing or emotional appeals. When the prompt explicitly included a requirement to check non-Python files, the agent succeeded regardless of other role-based additions. The author noted that none of the tested prompt variations resolved a specific bug related to negative number division in financial calculations.
Separately, Red Soft and Passion Tech launched the Molniya messenger for Android in the RuStore. The application is designed to replace foreign services by integrating social media, video streaming, payments, and cross-border trade features. Users are charged 750 RUB to download the messenger from the store.
