Project management has fundamentally shifted with LLMs. The author's time distribution changed from 70% coding, 15% meetings, 15% testing to 30% prompting agents, 10% meetings, 30% testing, and 30% extra time for other work. This radical shift requires new optimizations.

The first technique is mapping out work thoroughly beforehand. Unclear tasks cause agents to stop repeatedly for clarification. The author clears ambiguities by thinking through tasks or discussing with an LLM, then has the agent present ambiguities in an HTML report for review before autonomous execution.

The /goal command acts as a hook forcing agents to reflect on completion. Without it, models like Opus 5 can be "lazy" and stop prematurely. This command ensures agents work until all requirements are met.

Testing has become the new bottleneck. The author automates testing using Playwright MCP with Claude Code and Codex agents, which spin up localhost servers and test in Chrome. Explicit success criteria mapped beforehand let agents self-verify, catching issues like 404 errors automatically.