xhs-pharma-social-listening

XHS Pharma Social Listening

Overview

Use this skill to run an end-to-end Xiaohongshu social listening workflow for foreign pharma employee pain points. The main agent owns orchestration, scoring, analysis, and final reporting. Delegate Xiaohongshu collection to a collector subagent when subagent tools are available. Use Agent Reach/OpenCLI for access checks, seed discovery, spot validation, and note-level fill-in. For large runs or when OpenCLI throughput is insufficient, prefer a batch crawler path, with MediaCrawler as the first candidate. See references/batch-collection.md before changing collection strategy.

Workflow

1. Check platform access:
   • Run agent-reach doctor --json.
   • Confirm xiaohongshu.status is ok and active backend is OpenCLI or another working backend.
2. Select collection path:
   • Delegate Xiaohongshu information collection to a subagent when multi_agent_v1.spawn_agent or an equivalent subagent tool is available.
   • Keep one active XHS collection subagent per logged-in account/session unless the user explicitly asks for parallel collection; platform rate limits and captcha risk matter more than raw parallelism.
   • If the task is small or only needs validation, run scripts/collect_xhs_pharma.py.
   • If the user reports OpenCLI is slow, requests >1000 items, or asks for recurring social listening, read references/batch-collection.md and run a MediaCrawler POC first.
   • For large runs, target at least 1,000 independent evidence items: deduplicated notes plus deduplicated comments. Report search rows, note-detail rows, and comment rows separately; a detail row enriches its note and is not another independent item.
   • Use a broad-first query matrix. Start with single company, alias, and pharma-specific role terms, then add paired precision probes. Do not run cross-industry work or pain terms as standalone searches, including 市场部, 报销, 背调, 外企, 裁员, 离职, 合规, 薪资, and KPI; these must be paired with a pharma context anchor from references/query-matrix.md.
   • Cover company names, blackwords/aliases, roles, pain terms, compliance terms, AI/medical affairs terms.
   • Do not narrow or abandon the broad-first query matrix merely because MediaCrawler returns CAPTCHA/461. First distinguish query semantics from backend execution risk. If Agent Reach/OpenCLI can still search/read the same query, treat the issue as MediaCrawler-path or session cooldown risk, not as evidence that the keyword strategy is invalid.
   • Before any MediaCrawler batch run after a prior CAPTCHA/461, run an Agent Reach/OpenCLI smoke on the same or adjacent keyword: opencli xiaohongshu search "<query>" --limit 5 -f json, then read one note and comments. If OpenCLI succeeds but MediaCrawler fails, switch the collection backend to Agent Reach/OpenCLI and cool down MediaCrawler; do not continue retrying MediaCrawler in the same turn.
3. Analyze:
   • Run scripts/analyze_xhs_pharma.py on the collection directory.
   • Produce report.md, topic_scores.csv, evidence_notes.csv, evidence_comments.csv, and dataset_summary.json.
4. Report carefully:
   • Treat output as sampled social listening, not platform-total statistics.
   • Mark unverified items: author identity, policy authenticity, company-specific generalization, and whether a pain point is industry-wide.