Assistant for user stories

[SommerEngineering]

An assistant for user stories (see Wikipedia) designed to simplify the process of formulating requirements. To maximize benefits, the assistant should not generate just a single story—because for a single story, the task would simply become an aid for phrasing or rephrasing a particular statement. Instead, the approach could look like this:

• The input is a loose collection of bullet-pointed requirements from various stakeholders or roles.
• Later, this input can also be an audio recording, which we convert into a transcript.
• Additionally, we predefine possible roles.

Using this information, the LLM can then generate N user stories. However, many models will struggle to achieve this in one step. Therefore, we must implement a multistep process:

• System prompt A: An LLM extracts requirements from the bullet-point list or transcript, including implicit ones. For example, in order to use a multi-user web application, the user must first register and then log in later.
  • Idea: feed the LLM just one bullet point at a time, rather than processing the entire list at once.
  • For a transcript, we will need to implement preprocessing and potentially use another agent for that.
• System prompt B: An LLM assigns one of the N predefined roles to each identified requirement.
• System prompt C: An LLM takes the requirement and associated role and formulates the user story in the selected language.

Finally, we deliver to the user a Markdown-formatted list containing all generated user stories.