MANDATORY entry point for ALL research tasks. Orchestrates gatherer agents — brain decides agent count, languages, iterations, and depth at runtime based on the actual question. Default languages: EN + ZH + ZH-TW. Triggers: any research request, 'research X', 'what's the best', 'compare X vs Y', 'how should I approach', 'evaluate X', 'deep research', 'comprehensive analysis', 'team research' (live supervised agent-team venue). ALL research goes through lead-researcher first — never spawn gatherer agents directly.
Use when the user wants to design, redesign, shape, critique, audit, polish, clarify, distill, harden, optimize, adapt, animate, colorize, extract, or otherwise improve a frontend interface. Covers websites, landing pages, dashboards, product UI, app shells, components, forms, settings, onboarding, and empty states. Handles UX review, visual hierarchy, information architecture, cognitive load, accessibility, performance, responsive behavior, theming, anti-patterns, typography, fonts, spacing, layout, alignment, color, motion, micro-interactions, UX copy, error states, edge cases, i18n, and reusable design systems or tokens. Also use for bland designs that need to become bolder or more delightful, loud designs that should become quieter, live browser iteration on UI elements, or ambitious visual effects that should feel technically extraordinary. Not for backend-only or non-UI tasks.
Visual design verification loop for matching a goal PNG to a live render. Use when the user shares a design mock / goal image and asks to match the UI to it, asks to "iterate on this design" or "match this 1:1", or calls out visual drift in a previous UI iteration. Provides histogram color sampling, shape inspection (radius/border/shadow/bg-mode/fill-vs-context), computed-style checking on the live render, snap, and side-by-side diff. After every UI change the assistant MUST snap the live render and Read the snap with the Read tool before reporting the change done — skipping that is the
Deep internet gathering — the search-fetch loop, query templates, multi-source/GitHub patterns, and retrieval tools. ONE job: collect raw sources and data from the internet on an assigned topic/language. Run standalone to deep-gather a topic, or as the GATHER-DATA step of lead-researcher's research loop. Does NOT reason, steer, or synthesize — that is lead-researcher's brain. Triggers: 'deep gather', 'gather sources/data on X', 'collect everything about Y'.
Search GitHub for repos, code, and usage examples using gh CLI. Capabilities: repo discovery, code search, finding library usage patterns, issue/PR search. Actions: search, find, discover repos/code/examples. Keywords: gh, github, search repos, search code, find examples, how to use library, stars, language filter. Use when: finding repositories, searching code patterns, discovering how libraries are used, exploring open source.
ALWAYS invoke this skill when the user asks about Claude Code features, configuration, extensibility, or project setup. ALWAYS invoke when the user wants to create, improve, or refine a Claude Agent Skill — including 'make this a skill', 'turn this into a skill', 'create a skill for X', or asks about SKILL.md authoring, frontmatter, description optimization, progressive disclosure, or triggering strategy. Modules: CLI tool (setup, slash commands, MCP servers, hooks, plugins, CI/CD), extensibility (agents, skills, output styles), CLAUDE.md (project instructions, optimization). Keywords: Claude Code, skill, SKILL.md, agent, hook, plugin, MCP, CLAUDE.md, skill architecture, description optimization, progressive disclosure. For skill body content quality (soul, tensions, mental models), also invoke prompt-architect. Do NOT use for general prompt engineering without a Claude Code context.
Live collaborative brainstorm. Spawns named teammates (agent teams) that BUILD ON EACH OTHER'S ideas directly via SendMessage — generative cross-pollination ('yes, and…'), not one agent listing ideas alone and not an adversarial debate. Heuristic: use teammates when the agents need to talk to each other (build on ideas, hand-offs, debate); use parallel subagents when they don't (independent fan-out). Keywords: council, collaborative brainstorm, brainstorm together, ideate together, build on ideas, expand ideas, expand on this, riff together, group ideation, how might we, what could we build, team brainstorm. Use when: user types /council, says 'brainstorm together', 'expand on this idea', 'ideate as a team', or wants an open question explored by diverse perspectives that compound each other's ideas live. Requires the MAIN session (a subagent cannot spawn teammates) and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Do NOT use for: a quick idea list (use solo agents), deciding between options (use /debate), mechanica
Live multi-agent debate. Spawns named teammates (agent teams) that argue WITH EACH OTHER directly via SendMessage — a real, persistent debate, not one agent simulating both sides in a single reply. Heuristic: use teammates when the agents need to talk to each other (debate, collaborative brainstorm, hand-offs); use parallel subagents when they don't (independent fan-out). Keywords: debate, debate this, argue both sides, pressure-test, pressure test, steelman, devil's advocate, red team, red-team, challenge this, dialectic, adversarial review, two advocates, for and against, attack this idea. Use when: user types /debate, says 'debate this', 'argue both sides', 'pressure-test this decision', or wants a high-stakes call stress-tested by opposing positions that engage each other live. Requires the MAIN session (a subagent cannot spawn teammates) and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Do NOT use for: simple questions, mechanical tasks, or when one analysis suffices.