Prompting and Agent Roles

Start the backend and one client, then describe the outcome, relevant files or systems, and constraints. The root agent can answer directly, use tools, or delegate bounded work to a specialist role.

Examples:

• “Explain the failure in this log and suggest the smallest fix.”
• “Review these files for authentication regressions; do not edit them.”
• “Implement the issue, run the focused tests, and show me the diff.”
• “Research this question and cite primary sources.”

Specific constraints beat inferred preferences. State when a command must not run, when files are read-only, or when external communication requires a draft first.

Skills add task-specific instructions when discovered and selected. Roles provide specialist prompts and allowed capabilities. Topic classification can help choose a role, but phrasing does not guarantee a particular role or tool call. Permission policy remains authoritative: ALLOW runs, DENY blocks, and ASK requires an attended approval decision.

For reproducible work, include acceptance criteria and ask for verification that measures behavior rather than a fixed implementation detail.