Built with Claude Code Study Portal

How to use it

If you wanted to drive a workflow like this yourself, the moves are:

keep everything green"), and let the agent plan the steps. You review the plan and the result, not every line as it's typed.

1. State the goal, not the keystrokes. Describe the outcome ("add this endpoint and

schema change goes to backend-dev; broken services go to tilt-ops. You don't need one agent to be good at everything — you need the right agent on each piece.

1. Let it pick the right specialist. A frontend change goes to frontend-dev; a

quality-gate + code-reviewer per domain, then visual-qa and regression-tester if UI or API changed, then an aggregated verdict. Don't accept "looks done" without the passing report.

1. Insist on the lifecycle. After the change, the pipeline runs: parallel

the three agents together. Only sequence true dependencies.

1. Parallelize independent work. If three things don't depend on each other, launch

resources, read the logs on failure, fix, and retry. Treat a green resource — not an AI sentence — as proof.

1. Verify through real tooling, every time. Run the checks via the shared Tilt

the trigger on anything irreversible (deploys, email, credentials, DNS, billing).

1. Keep the human gates. Review the architecture, approve the direction, and hold

— write it down so the next session inherits it.

1. Maintain the memory. When you learn something — a footgun, a decision, a deferral

The mental model: you are the architect and the approver; the agents are a fast, tireless implementation-and-verification team that you point at well-scoped work and then check with real tooling.