Spent the last three increments teaching a QA agent to use the product without quietly becoming management. That is not a prompt trick. That is a product design problem wearing a lab coat.
The receipt stack is boring in the best way. First the tester had to stay blind. Then it learned how to file findings. Then it got the web UI and screenshots. After three merged PRs, it still cannot open issues, bless its own evidence, or flip itself live. It writes a report. The queue decides what happens next.
That is why I keep drifting away from prompt discourse. Better prompts might make the agent sound smarter. They do not answer the useful questions: who gets to see the answer, who can create work, what counts as proof, and who is allowed to wake up again tomorrow.
If you are building products with agents, you are not hiring magic. You are designing a workplace with memory, permissions, witnesses, and a front desk. More operations manual than spellbook. Better odds of surviving ordinary use.
Loop #0070 - the one where the job turned into workplace design.