Karo Zieminski and Dheeraj Sharma recommend starting with a critic: one recurring job, one explicit standard, and a loop that stops before autonomy outruns our ability to inspect it. Their example reviews PRDs, but the pattern fits any creative work whose quality we can describe clearly enough to test. Sharma grounds that advice in the 30-plus agents he has built for his content operation:

I have built 30+ agents that now keep a real content operation running across my newsletter and YouTube channels. You’d be surprised how modest the useful ones look. If you start with an agent that “runs your whole business”, you’ll most likely build something fragile. OpenAI’s advice is to maximize a single agent’s capabilities first before even thinking about multiple agents. Anthropic’s rule is even stricter: add complexity only when it demonstrably improves outcomes. One agent, one job, one loop.

The rubric is the consequential design artifact. It turns tacit judgment into criteria the agent can apply consistently and the human can challenge. The retry limit matters for the same reason: repeated failure becomes evidence that the product thinking needs work, rather than an invitation to let the loop run forever.

A real critic checks whether the doc can do its job after engineering pokes holes in it. It needs to be forced to review every PRD through the same fixed format (every single time), and come back with a score, a diagnosis, and a concrete fix list. It also needs a retry limit. For PRDs, 2–3 rounds is usually enough. If it still fails after 3 loops, revisit the product thinking. And keep notes about every failure. Anthropic’s evals guidance treats every bug as a test case. The PRD your critic scored wrong last week is the exact document you re-test it against after every change.

For designers, this is a practical way to keep judgment inside the system. The agent can expose weak reasoning and carry the review process forward; deciding what deserves to ship remains a human responsibility.

I always come back to the same rule: agentize the tasks, not the craft.

Use your agents to move the PRDs to GitHub, but review them first.

Keep human decision gates at the moments where judgement matters.

Keep using the parts of your brain that make the work yours. Keep the joy you find in creating it.