Research knowledge orchestrator

Research & Knowledge Orchestrator

The single entry skill for research and knowledge work. It locates the task on the question type × evidence depth map and delegates to one of 12 specialist spokes. The cross-cutting discipline every spoke shares — pick the lightest evidence lane that answers the question, label every claim by provenance, and escalate only when synthesis demands it — lives in research-knowledge-core; read it before promising coverage or mixing sources.

Routing map (intent → spoke)

Operate a research pass (start here when the lane is unclear)

• "Research this", "compare", "what's the latest", recurring lookup → research-ops (operator wrapper; chooses the lane below)

Current-web research

• Fast discovery / web · code · company · people lookup → exa-search
• Thorough, cited, multi-source report ("deep dive", "current state of") → deep-research

Academic & scientific literature

• Find · screen · synthesize · cite a body of literature → scientific-thinking-literature-review
• Judge a paper / proposal / methods section / evidence quality → scientific-thinking-scholar-evaluation
• Biomedical literature, MeSH, PMID, E-utilities → scientific-db-pubmed-database
• Patents & trademarks, official IP records → scientific-db-uspto-database
• Genomic database queries, sequence lookup, enrichment → scientific-pkg-gget

Understand a codebase

• Map an unfamiliar repo → onboarding guide + starter CLAUDE.md → codebase-onboarding
• Author a step-by-step .tour walkthrough (onboarding / PR / RCA) → code-tour

Reference & memory

• Up-to-date library/framework docs (named framework, API, setup) → documentation-lookup
• Persist project context across sessions; resume where you left off → ck

Folded spokes (content extraction, scraping, transcripts, monitoring)

Routable spokes folded into this cluster. They cover the acquisition and distillation lanes — getting raw content out of the web/feeds/recordings and turning it into structured, summarized evidence — feeding the research lanes above.

General research & content distillation

• Three-mode research (quick/standard/extensive) + content extraction; 240+ Fabric patterns → research
• Current open-web search with source extraction and evidence gathering → web-search
• Apply a named Fabric pattern (extract wisdom, summarize, threat model, etc.) to content → fabric
• Summarize/transcribe a URL, podcast, or local file (text + transcript fallback) → summarize

Parse & extract structured content

• Parse URLs, files, videos, PDFs, articles to structured JSON (entities, transcripts, batch) → parser
• Fetch + summarize a YouTube video's transcript (proxy-backed for cloud IP blocks) → youtube-transcript

Scrape the web at scale

• Social-media / e-commerce / business-data scraping via Apify actors (Twitter, IG, LinkedIn, TikTok, Maps, Amazon) → apify
• Progressive, tiered URL scraping via Bright Data → brightdata

Meetings & feeds monitoring

• Analyze meeting transcripts/recordings for communication patterns and actionable feedback → meeting-insights-analyzer
• Interact with Fireflies meeting data and the Fireflies API (via Membrane) → fireflies
• Monitor blogs and RSS/Atom feeds for updates (blogwatcher CLI) → blogwatcher

Standard Operating Flow

1. Classify the ask: which question type (current fact · comparison · literature · IP · genomic · codebase · API reference · resume context) and which evidence depth.
2. Take the lightest useful lane first — local/docs/memory before web, exa-search before deep-research, scoping review before systematic. The model and escalation ladder are in research-knowledge-core.
3. Delegate to the spoke(s). Multi-step asks fan out in evidence order (e.g. "review the literature and rate the key paper" → scientific-thinking-literature-review → scientific-thinking-scholar-evaluation).
4. Return: chosen spoke(s), the evidence lane used, claims labeled by provenance (sourced fact / supplied context / inference / recommendation), dates on freshness-sensitive answers, and the next action.

Guardrails

See research-knowledge-core. In short: evidence-tier discipline — never answer a current question from stale memory when a fresh search is cheap; never mix inference into sourced facts without labeling it; never spin up a heavyweight research pass when local code, docs, or ck memory already hold the answer; always date freshness-sensitive claims and name your sources. The cluster's value is trustworthy, traceable answers — don't trade that for speed.

Boundaries

• "Research this" — which spoke? Default a bare, lane-unclear "research this" to research-ops, which picks the rung. Skip it and go direct when the lane is already obvious: deep-research for a thorough multi-source cited report on a general/web topic; the scientific-* spokes (scientific-thinking-literature-review, -scholar-evaluation, scientific-db-pubmed-database, -uspto-database, scientific-pkg-gget) when the subject is academic / biomedical / IP / genomic and needs scholarly rigor or a citable database.
• Codebase, not the web. codebase-onboarding and code-tour operate on this repo's own source — mapping it and authoring .tour walkthroughs. They are not web research; for external topics use deep-research / exa-search instead.

Picked-up spokes (knowledge-base authoring & scientific computation)

Vetted additions from the antigravity-awesome-skills library (MIT). They extend two lanes the cluster was thin on: personal knowledge-base authoring (Obsidian vaults) and exact scientific computation, plus one more content-extraction tool.

Knowledge base authoring (Obsidian)

• Write/edit Obsidian Flavored Markdown — wikilinks, embeds, callouts, properties → obsidian-markdown
• Build .base database-views (filters, formulas, table/card/list) over notes → obsidian-bases
• Author/edit .canvas visual maps, mind maps, knowledge graphs (JSON Canvas 1.0) → json-canvas
• Shell-driven vault ops + plugin/theme dev against a running Obsidian → obsidian-cli

Scientific computation (turn sourced data into exact results)

• Symbolic math: solve equations, calculus, simplification, closed-form derivations → sympy
• Astronomy: coordinates, units, FITS I/O, WCS, cosmology, precise time → astropy

Content extraction

• Token-lean clean-markdown extraction from a webpage via the Defuddle CLI → defuddle (use over a raw page fetch when reading docs/articles/blog posts; complements parser/web-search)

Loading spokes on demand

To keep CLI startup context lean, this cluster's spokes are not separately registered as skills — only this orchestrator and its *-core are enumerated. When you route to a spoke named above, load it on demand by reading its file:

~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md (or skills/<spoke-name>/SKILL.md inside the skill-clusters repo).