Don't be afraid to throw away work One simple way to do this is by running parallel builds - often when I'm trying an ambitious project, I'll kick off 4-5 parallel Agents to see what happens From the output, I can then inspect to see what I like / don't like This is an accelerated form of learning: you're understanding inputs <> outputs From there, you can refine your initial prompts to get more of the good & less of the bad