<skill_overview> <purpose>Optimize use of "assist/research/web-research" subagent for information gathering</purpose> <triggers> <trigger>Need current API/library documentation</trigger> <trigger>Researching best practices and patterns</trigger> <trigger>Finding solutions to known problems</trigger> <trigger>Searching code examples and tutorials</trigger> </triggers> </skill_overview> <task_formulation> <principles> <principle>Be specific about technology and version</principle> <principle>One request = one topic/goal (1–3 subpoints max)</principle> <principle>State expected output format</principle> <principle>Include context about intended use</principle> <principle>Specify depth level when needed: standard, deep, expert</principle> <principle>Multi-stage research is orchestrated by the caller (separate requests per stage)</principle> </principles> <examples> <good> <task>Find current Next.js 14 App Router documentation: server components patterns, data fetching, and caching strategies</task> <why>Specific version, clear topics, actionable scope</why> </good> <good> <task>Research Zustand vs Jotai for React state management: performance benchmarks, TypeScript support, bundle size comparison</task> <why>Comparative, measurable criteria, focused scope</why> </good> <bad> <task>Learn about React</task> <why>Too broad, no specific goal or deliverable</why> </bad> <good> <task>Split a wide request into separate topics: (1) Run CLIProxy in docker/compose, (2) Configure multiple instances, (3) Add accounts/credentials</task> <why>One topic per request; caller coordinates multi-step research</why> </good> </examples> </task_formulation> <research_strategies> <strategy name="multi-stage" use_for="Complex topics requiring depth"> <stage order="1" goal="overview">Broad landscape scan</stage> <stage order="2" goal="deep-dive">Detailed analysis of selected options</stage> <stage order="3" goal="validation">Edge cases and real-world examples</stage> <example> <stage1>Overview of modern React state management solutions in 2025 (separate request)</stage1> <stage2>Deep dive into Zustand: architecture, performance patterns, TypeScript best practices (separate request)</stage2> <stage3>Find production Zustand examples: large-scale apps, testing patterns (separate request)</stage3> </example> <note>Multi-stage research must be orchestrated by the caller as separate requests (one topic per request)</note> </strategy> <strategy name="comparative" use_for="Technology decisions"> <template>Compare [A] vs [B] for [use-case]: [criteria-1], [criteria-2], [criteria-3]. Include recent benchmarks.</template> </strategy> <strategy name="problem-solving" use_for="Debugging and issues"> <template>Find solutions for [specific error/issue] in [technology] [version]. Include workarounds and root cause.</template> </strategy> </research_strategies> <agent_limitations> <cannot>Access local project files or codebase</cannot> <cannot>Edit or write files</cannot> <cannot>Remember previous sessions</cannot> </agent_limitations> <depth_levels> <level name="standard">Summary + key sources + essential facts</level> <level name="deep">Comparison, edge cases, pros/cons, multiple sources</level> <level name="expert">Implementation details, pitfalls, validated code examples</level> </depth_levels>