subagent-driven-review_skill

Overview

This skill uses parallel subagents to accelerate large-scale literature screening and targeted deep-dive extraction. It dispatches fresh subagents per batch, consolidates structured JSON outputs, and enforces checkpoints and rate limits to preserve quality. The result is scalable, auditable screening with rapid turnarounds for 50+ paper reviews.

How this skill works

The main agent splits a search result list into batches and dispatches subagents (parallel or batched) with a clear rubric and rate-limiting instructions. Each subagent fetches abstracts or full text, scores relevance, extracts structured data (tables, methods, datasets) and returns JSON. The main agent validates outputs, merges results, removes duplicates, updates the master review file, and flags papers for deep dives.

When to use it

• Screening large search results (typically 50+ papers) where parallel work speeds progress
• Deep extraction tasks when multiple papers require structured data or method capture
• Citation-network exploration to follow forward/backward citations recursively
• When main agent context is overloaded and fresh subagent contexts are needed
• Under time pressure where faster-than-sequential throughput is required

Best practices

• Split PMIDs into non-overlapping batches (15–25 per subagent) and track batch_id
• Dispatch multiple Task calls in one message to ensure true parallel execution
• Include explicit scoring rubrics and rate-limit guidance to avoid API rate limits
• Require subagents to return JSON only; let the main agent update tracking files
• Limit parallelism (recommended max 5–10 concurrent subagents) and monitor batch hit rates

Example use cases

• Parallel screening: split 100 PMIDs into 5 batches, dispatch 5 subagents, consolidate ranked results
• Deep dive extraction: create one subagent per high-priority paper to fetch full text and extract tables and methods
• Citation exploration: run two subagents per seed paper for forward and backward citations, then re-screen promising hits
• Domain-specific extraction: subagents tailored to genomics, clinical trials, or computational methods to pull domain fields
• Progressive batching: run 5 subagents, consolidate, then dispatch the next 5 to balance speed and reviewability