research-deep

Place an outline.yaml with an items list in your working directory and run /research-deep. The Skill batches items and launches background web-search agents in parallel, asking for approval before each new batch.

Rerun /research-deep in the same directory. The Skill checks the output directory for completed JSON files and automatically skips items that already have results, continuing only with unfinished items.

It requires an outline.yaml containing the topic, items list, and execution config such as output_dir and items_per_agent, plus a fields.yaml defining the output schema. Both are typically produced by the preceding outline generation phase.

Each agent writes structured JSON following fields.yaml definitions, marks uncertain values with [uncertain], and lists them in an uncertain array. A validation script then checks complete field coverage before the task is considered done.

Each batch requires user approval before the next starts, so the workflow is not fully unattended. Output quality also depends on web search availability, and uncertain fields must be reviewed manually afterward.