Suggest improvements to the analysis pipeline prompt after evals

When having enough evals, we should suggest prompts that might correct the mistakes the eval surfaced, allowing to re-run an eval and compare results to help improving it.