Deep, multi-source, fact-checked research with citations.
Uses
APIs to connect
This skill needs keys for these services. The link takes you where to get each one:
What it does
Spawns a sub-agent pipeline (scoper → source-hunter → deep-reader(s) → synth → judge) with a cite-everything protocol, producing report.md with sources and confidence ratings. 5 modes: quick (a fast fact in ~10s, single pass) / lite / standard / heavy / ultra. Search isn't just the web — best-in-class engines are wired in: Exa (neural/semantic), Tavily (LLM-native with a ready answer), Perplexity (grounding), plus academic primary sources (arXiv/GitHub via firecrawl) on deep modes; all hits collapse via canonical dedup (arxiv/DOI variants → one source). Verification is hardened with an adversarial pass: key claims are re-checked by an independent search TO REFUTE, with a quality bar (refuted only if the counter-source clears the same bar).
How to use
- 1Say “research X” or
/research <topic>. - 2The scoper picks a mode (asks to confirm for heavy/ultra).
- 3Agents gather and cross-check sources.
- 4You get a cited report with confidence levels.
When it triggers
Example
/research the RU EV-charging market 2026 — gathers sources, fact-checks, returns a cited report.
Updates
The external reviewer model no longer runs at a lowered temperature: for reasoning models Google explicitly recommends keeping the default, otherwise the risk of looping and degraded analysis goes up. The value is aligned across all copies, and the test that guarded the old proxy implementation is re-pinned to what actually matters.
Privacy barrier before external models: topic, report and sources pass a fail-closed scrub (mask secrets/infra, block sensitive data) before egress to GPT-5/Gemini for a cross-model second opinion. Privacy model unified with plan-panel and finalize.
Depth + speed upgrade: (1) a new quick mode — a fast cited fact in ~10s, single pass, auto-triggered (fresh web fact with no explicit request); (2) three best-in-class search engines — Exa (neural), Tavily (LLM-native), Perplexity (grounding) — under one contract with fail-open, privacy scrub and an atomic budget guard; canonical dedup collapses one source found by several engines (arxiv/DOI variants too); (3) an academic layer — arXiv/GitHub primary sources on deep modes; (4) adversarial-verify — key claims are re-checked by an independent search TO REFUTE with a quality bar. All under per-mode toggles and budget caps.
Added to the RedSkills catalog.
Liked this skill? New breakdowns and updates land in the channel.