Turn a research question into an auditable evidence set and an answer-first report.

Clarify the topic, audience, geography, language, time window, platforms, and desired deliverable from the request. Infer low-risk omissions and record them. Never invent access to a platform: list unavailable or excluded sources in the report.

Create a resumable run:

python3 <skill-dir>/scripts/scholar.py init --topic "<topic>" --output <workspace>/research-runs

Read references/method.md before collecting. Read only the platform sections needed from references/platforms.md. For repeat or scheduled monitoring, also read references/monitoring.md.

Prefer, in order:

Do not bypass authentication, CAPTCHAs, paywalls, access controls, or platform safeguards. Do not collect private profiles or unnecessary personal data. Stop on explicit denial, repeated throttling, or a platform instruction prohibiting the attempted access method; record the gap and continue with other sources.

Write one JSON object per observed item to . Preserve the source URL and observed facts; keep interpretation in or the final report. Use the helper where practical:

python3 <skill-dir>/scripts/scholar.py add <run>/items.jsonl \
  --platform <platform> --url <url> --title <title> \
  --published-at <ISO-8601> --observed-at <ISO-8601> \
  --metrics likes=10 comments=2 views=300 --text <summary>

Engagement values are platform-specific snapshots, not comparable audience counts. Store query-result pages only as discovery evidence; open the underlying source before using it to support a factual claim.

Periodically generate the working report:

python3 <skill-dir>/scripts/scholar.py report <run>

Use trend scores to prioritize reading, never as proof of importance. Synthesize recurring themes, emerging signals, disagreements, audience language, and actionable opportunities. Distinguish observed evidence, cross-source inference, verified claims, and open uncertainty.

Stop when the requested deadline or cap is reached, every in-scope platform is covered or documented as unavailable, and the latest collection round adds no material theme or changes no conclusion. For open-ended requests, default to two collection passes and at least two independent sources for each major conclusion; extend only while new evidence materially changes the answer.

Update with an executive answer, scope and method, strongest trends, platform differences, evidence links, opportunities, caveats, and uncovered gaps. Cite the direct item or authoritative source beside each claim. State exact collection dates and do not describe a trend as current outside the observed window.