Prompting agents to find optimization opportunities and bugs in Elixir codebases

I like to find performance optimizations and hidden bugs in Elixir codebases with coding agents.
If I just ask them directly to find those, they are not that successful, but recently I found two ways that are more fruitful.

Core Erlang and BEAM bytecode inspection:
If I ask my agent to inspect the Core Erlang and BEAM bytecode of a module, it finds much more insightful details and higher quality performance patches.

One of my prompts looks like this:

Further refactor ModuleName.function towards the Pareto frontier of performance, simplicity, elegance.
First, always check the Core Erlang & BEAM bytecode. Then do the refactor, then check Core Erlang & BEAM bytecode again so you continuously gain performance insights from them.

If there are low hanging performance improvement opportunities, the above prompt can find those usually.

Lean 4 formalization:

I would do something like this:

/goal Fully formalize module N of this Elixir codebase in a new formalization/ dir with the Lean 4 theorem prover.

Then it goes to work for a long time (it probably never finishes), but while the main thread is running, I’m using the /btw command to ask questions about the code based on the formalization. This can find very interesting things about the code and it seems like the optimizations/bugs/security issues it can find this way are numerous.

My Conclusion:
Maybe these are just different ways to force an agent to do work and to look at code from different perspectives, but I found that these work well enough to find surprisingly good issues in established codebases.