← ClaudeAtlas

scan-researchlisted

Scans a topic in one pass and returns one sourced Markdown report. It writes a four-line question analysis (the decision, the question type, the presuppositions, the useless answer), derives three or four sub-questions from the decision, and launches one collection agent per sub-question in parallel, each on a budget of eight tool calls with one query spent on what would make its answer wrong. Before delivery, a verification agent in a fresh context checks up to five Answer sentences against their sources, and the main session rewrites what it flags. In the report, every claim has a full URL, a publication date and a source grade, every key figure has a quote from its page, and the Answer states the mechanism behind it. Web-first, plus any documents the user supplies; academic sources come in only where a sub-question turns on effectiveness, causality or a comparison of methods. Use it when the request wants coverage of a subject with its sources attached but does not warrant a verified evidence ledger: 「ざっと調
tomkd555/scan-research · ★ 0 · Data & Documents · score 75
Install: claude install-skill tomkd555/scan-research
# scan-research A scan is the step between one search and a research-team run. Spend its tokens on searching and on checking the sentences the Answer rests on, and none on re-reading. The collectors search and read the pages. The main session reads only their evidence files, the verdicts and the agents' summaries, so its context stays small for every turn that re-reads it. The one exception is the fallback under Errors. Its safeguards are a fixed collection budget, a grade on every source, a quote from the page on every key figure, and one verifier that checks up to five Answer sentences without seeing the report. research-team adds a script that matches every quote against a cached copy of its page, verification of up to 10 sentences (STANDARD) or 25 (DEEP), and a script check of the report and its translation. Pause for the user only at the scope question in Step 1 and the hand-over to research-team. Before each step, say in one line what it does; when a wait ends, say in one line what came back. While agents run, do the work that does not depend on them. When nothing is left, end the turn with one line on what is running; their notifications resume the scan. **Language.** Section headings, table headers and the confidence levels are English. The prose under them follows the user's standing convention for deliverables. ## Step 1 Frame Draft the report's header first and keep it as written. - Purpose: what this scan is for, in one line. - As-of date `{AS_OF}`: today's