The Operating Loop

• Treat the repository as a brain by creating a specific structure: /context for business knowledge, /customers for feedback and objections, and /routines for recurring tasks.6:40
• Encode the operating system in three root markdown files: claude.md defines how the AI works, roadmap.md sets the current focus and out-of-scope boundaries, and review.md establishes a rubric for judging quality before shipping.7:28
• Adopt a "measure twice, cut once" discipline by requiring the model to generate a plan and identify verification steps before it touches any code.2:28

Scaling and Oversight

• Maximize output quality by assigning small, clear tickets rather than vague goals, ensuring the AI can see a concrete finish line without needing to guess intent.
• Use "eyes" — desktop previews and manual inspection — to catch trust or clarity issues, such as a landing page failing to reassure a visitor about waitlist expectations.3:23
• Separate work into isolated sessions or worktrees to run parallel assignments, keeping bug fixes, feature work, and demo scripting tasks distinct and manageable.35:08

Risk Boundaries

• Enforce a layered review process: humans inspect the code diff for intent, while Claude critiques its own work against the review.md standards for bugs and security risks.26:56
• Define clear permission tiers for safety: keep production deploys, customer data handling, and billing decisions strictly "human-owned" while allowing the AI freedom within feature branches.39:15