deep-research · hkustdial — Claude Code skill: literature survey with verified

What deep-research does

The skill will not start searching until a brief exists: topic, two or three answerable research questions, the angle, and who the report is for. A vague request gets turned into questions first. From there it opens three to five search perspectives, among them the mainstream line of work, its critics, nearby disciplines, methodology and deployment. Each perspective searches broad, then narrows using terms the first pass turned up. If the environment can spawn sub-agents, the perspectives run side by side; otherwise you run them one after another, same method.

Every candidate reference goes through a two-step check, and a paper that cannot be confirmed stays out of the corpus. Findings are then sorted into a MECE taxonomy, and sections are written one at a time, each citing only the works assigned to it. Sources get compared within a sentence. Where studies disagree, the report explains under what conditions, rather than splitting the difference. A self-adversarial pass follows: three retrospective questions plus a check for an omitted perspective.

Six internal gates watch the work — angle, coverage, citation, taxonomy, calibration, weaving. Failing citation or coverage sends you back for a targeted supplement rather than a restart. No retrieval, no run: the skill states that plainly instead of writing a survey from memory. The report follows a paper layout, from abstract and methodology to open problems and references.