Driving

Once your agent is wired, your job changes shape. The agent runs the method: it elicits, authors the chain, drives the build queue and keeps the referee green. You steer: you answer its questions, review what it captured, unblock builds and step in when something looks wrong. These pages are the craft of that supervision. Almost none of it is typing commands; most of it is knowing what to look at and what to say.

1. Steering what gets built

   The elicitation gate is where you shape the product. What the agent will ask, how to answer, and why "I do not know" is a legitimate answer.

2. Reviewing the tree

   The docs-review gate before the first build: open the viewer, read what your agent captured against what you meant, push back on what is missing.

3. When it goes wrong

   The agent checks its own work and raises anything it finds in plain language. What those findings sound like, and how to answer them.

4. Handing over

   Leave a session so the next one picks up cleanly, whether the next driver is you, a colleague, or a fresh agent session.

The rules your agent is following throughout are not hidden. They are the managed block rcf init wrote into CLAUDE.md and AGENTS.md, in your own repository, readable any time. The harness block publishes the same text on this site.