Demand it use everything it builds, end to end before ending its turn. So, the important first step in its plan is always the testing harness. Does it have one already, what would it need to add, etc. If it can run an end to end test then it can check its work, see it failed, loop to fix its failures, etc.
Unit tests, visual regression testing suite, integration testing suite, e2e testing suite etc.
If Nostr, that would include local relays. If streaming, local streaming servers. Etc
That and ask if to create review subagents that focus on documentation, architectural patterns, adversarial use, non-happy paths etc
If it does all that before it comes back to you, the result is more likely to be worth your time to review.
Katie Daviscourt: "Yeah $40,000 to $70,000 per daycare is what I was finding out in these two small little neighborhoods in the Seattle metro... and I was accused yesterday by the SEIU of conducting racist attacks on Americans... they sent out an email to all of the Washington state lawmakers last night saying that I was doing racist attacks on immigrants."
https://friends.ravergram.club/system/media_attachments/files/117/328/300/153/560/878/original/4ea979881c4d3d37.mp4