Moving from generating code to mapping code is already a major leap. But to break through the "feed data and read Markdown" plateau, you must transition from passive reading to active, structured prompting.

You can achieve this by moving from summarize to deconstruct. For example, standard agents default to generic summaries. You must force the agent to act like a senior engineer conducting a code review or a research lead. You can do this by prompting for architectural boundaries and friction points, not just descriptions, by asking "why", not "what".

You could also use tactical prompts for Code onboarding. An example prompt would be, "trace a single user request from the entry point API down to the database. List every module it touches and where mutations happen."

This is how I would do it if I were in your shoes, but there are certainly other methods that could be better suited for you. Either way, I hope this was at least useful in some way.