researchlisted

# Research ## Overview A four-mode research skill with anti-sycophancy safeguards. Usable by both humans and agents. The caller selects a mode (socratic, direct, deep, adversarial) or the skill auto-detects from the question's intent. **Core principle:** Every claim must cite a source. Unsourced assertions are labeled "unverified." The skill resists premature agreement through four safeguards that run silently across all modes. ## Modes ### Socratic Explore a problem space through guided questioning before researching. **Workflow:** 1. Ask clarifying questions **one at a time**, up to 5 rounds. Each question narrows scope, surfaces assumptions, or identifies what the requester actually needs. 2. After clarification (or if the requester says "just research it"), research using available tools. 3. Present findings with citations. Ask a follow-up question to go deeper if appropriate. 4. Iterate until the requester is satisfied. **Delivery depends on the requester.** When a human invokes the skill, ask each clarifying question through the AskUserQuestion interactive prompt, with concrete selectable options so the requester can pick rather than free-type. When an agent invokes the skill (agent-to-agent), ask in prose inside the reply message — agents cannot answer an interactive prompt. Detect the requester from context: a teammate or delegated subagent message is agent-to-agent; a direct user turn is human. **One question at a time.** Do not batch multiple clarifying qu