research-lean-linear

Provide a question and today's date; the skill decomposes it into search queries, retrieves web sources, distills claims, and writes a markdown brief with inline citation URLs. The final artifact is brief.json containing the report, citations, confidence, and gaps.

Each source is distilled with a relevance score and a credibility score from 0 to 1, weighing authorship, primary versus secondary status, and evidence shown. Claims are drawn only from the source content, and URLs are copied verbatim so every claim stays citable.

Run everything inline in one context by default, since it scores as high or higher at a fraction of the cost. Dispatch subagents only when combined inputs exceed your context budget, a step needs a different model or tool, or you need wall-clock parallelism.

Claims are marked unverified rather than independently fact-checked, so the brief sits at the cheap end of the quality/cost frontier. Unverified claims are never asserted in the report and are surfaced in the gaps field instead.

The pipeline writes queries.json, sources.json, distillate.json, and brief.json to the working directory. The final artifact is brief.json, containing report_markdown, a de-duplicated citations list, an overall confidence score, and gaps.