← ClaudeAtlas

deep-researchlisted

Deep multi-hop research protocol - decompose the question, run a multi-modal sweep (Exa neural and advanced search, Jina web plus arXiv/SSRN, Spider crawls of key sites, Playwright for JS-heavy pages), triage sources, triangulate every key claim across at least 2 independent sources, keep a contradiction log, critique completeness, loop until dry, and write a cited report to a file. Load for broad, contested or multi-part research questions and to decide when to launch the research-sweep workflow instead.
rustcircle/c101 · ★ 0 · Data & Documents · score 65
Install: claude install-skill rustcircle/c101
# Deep research Outputs (paths relative to your `out_dir`, normally `.stack/out/deep-researcher`): ``` <stamp>-<slug>.md the report (output-contract offload target) <stamp>-<slug>/plan.md scope card, sub-question tree, hypotheses, query log <stamp>-<slug>/sources.jsonl one source record per line (web-search-tools, section 5) <stamp>-<slug>/claims.md claim ledger (phase 4) <stamp>-<slug>/contradictions.md contradiction log (phase 5) <stamp>-<slug>/notes/S<n>.md per-source notes for anything read in full ``` `stamp` = `date +%Y%m%d-%H%M`. Keep ledgers in files, not in context: after each phase your context should hold the plan and counts, not page text. ## Phase 0. Route the job | Situation | Route | |---|---| | 3-5 sub-questions, about 25 sources or fewer | run this protocol yourself, delegating reads and fact-checks | | 6-12 sub-questions, each needing web + papers + site corpus | main thread only (no `<stack-context>` header): Workflow `research-sweep`, `depth: "standard"` (keeps 6 sub-questions; pass `maxSubQuestions: <n>` above that); as a subagent: run Phase 2 yourself with parallel Agent calls, or return the Workflow recommendation below as an Open doubt | | Landscape, market map, systematic review, "every X" | main thread only: Workflow `research-sweep`, `depth: "exhaustive"` (keeps 10; set `maxSubQuestions`, max 16); as a subagent: swarm-coordinator sharded by sub-question or source (only when `remaining_depth` >= 3), or the Workflow recomme