deep-research

A research harness that runs broad parallel web searches, validates findings across multiple sources, tracks confidence, and produces a cited Markdown report. It supports eleven research types including market, domain, technical, competitive, product, academic, person/org, financial, legal, trend, and community, and triggers on requests phrased formally ("competitive analysis", "literature review") or casually ("look into X", "what's the deal with Y"). It is built for Claude Code or similar agents and requires internet access through WebSearch and WebFetch.

Execution is organized into modes and depths. An interview mode scopes the work sequentially, a parallel research mode fans out 3 to 20 sub-agents per step with each owning one axis, and a synthesis mode reconciles conflicts using ultrathink before recommending. Depth is selected automatically: Quick runs scoping plus a couple of steps for narrow or time-sensitive questions, Standard runs the full Steps 1 through 5 by default, and Deep adds outline refinement and a critique pass for comprehensive or critical decisions. For well-scoped prompts it states assumptions in the report header and proceeds without a full interview, reserving the scope interview for genuinely vague prompts.

The skill is strict about evidence discipline. Web search is treated as its core capability, so it halts if WebSearch is unavailable. Every claim must cite a source URL, and critical claims such as market size or growth rates require two or more independent sources or are marked low confidence. Findings are written to the output file immediately after each step rather than batched, conflicts between sources are flagged explicitly, and output is prose-first, aiming for at least 80% full sentences with bullets reserved for true lists. It distinguishes sourced facts from analytical synthesis and admits gaps explicitly. Type-specific reference files and citation and parallel-search guides are loaded only at the step where they are needed. Use it whenever the deliverable is a thorough, sourced written analysis rather than a quick answer.