knowledge-work

Facilitate structured brainstorming using parallel sub-agents to explore idea spaces. Use for IDEATION/CONCEPTUAL WORK ONLY, NOT for implementation planning or task breakdown. Use when user wants to brainstorm, explore ideas, generate concepts, develop vision, or discover creative directions. Transforms vague ideas into practical, tangible expressions through 5 rounds of parallel agent analysis and refinement.

Web research via Gemini CLI with Google Search grounding. Generates markdown report. Claude MUST append analysis section after. Use for real-time data: news, docs, code examples, fact verification.

Search GitHub for real-world code examples and implementation patterns. Use when user wants to find code examples on GitHub, search GitHub repositories, discover how others implement features, learn library usage patterns, or research architectural approaches. Fetches top results with smart ranking (stars, recency, language), extracts factual data (imports, syntax patterns, metrics), and returns clean markdown for analysis and pattern identification.

Comprehensive web research via Parallel Search API. Use when user requests parallel search for deep multi-source research, technical analysis, learning new topics, current events, or comparative studies. Returns LLM-ready ranked URLs with extended excerpts (up to 30K chars). Single API call handles multiple query angles with automatic deduplication.

Fetch and analyze Readwise reading activity for any date range. Use when user asks about articles saved, highlights created, or most-highlighted content for today, yesterday, last week, last month, or custom date ranges (e.g., "show my Readwise activity from Jan 1-7"). Requires READWISE_API_TOKEN env var. Connects to Readwise Highlights API (v2) and Reader API (v3).

Download and aggregate web pages/docs into timestamped scratchpad files. Use when user asks to "concatenate all these resources", "get all these links", "checkout all these resources", or wants to gather fresh context from documentation URLs. All URLs from one prompt go into single file at docs/scratchpad/<timestamp>.md.

Batch-process web pages via headless Playwright browser, extract HTML, convert to markdown using Turndown, and save to timestamped scratchpad file. Use when user asks to "capture these pages as markdown", "save web content", "fetch and convert webpages", or needs clean markdown from HTML. All URLs from one prompt → single file at docs/web-captures/<timestamp>.md.