cowork-research

cowork-research

A Claude Cowork plugin that turns your assistant into a full-time research analyst — brand briefs, web audits, and meeting recaps that compound across every business decision you make.

Stop pasting links into ChatGPT. Your assistant already knows your business, your ICP, and your offer. Give it the right tools and it researches in your voice.

Built by Nuno Tavares for VCInc cohort students and anyone who wants research that accumulates instead of evaporating after every conversation.

---

Why this exists

In April 2026, Andrej Karpathy published LLM Wiki, a gist that crystallized the markdown-vault-as-second-brain pattern:

"Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase. You rarely ever write or edit the wiki manually."

Corey Ganim's framing applies the same idea to research specifically:

"Stop treating Cowork like a chatbot. Treat it like an employee. The first day on the job, you tell it about your business. After that, every research task is just delegation."

Same pattern, two voices. Both point at the same thing: a research workflow where your assistant already knows the context, has the right tools wired up, and writes outputs that match how YOU think. This plugin is the install + scaffold + wiring that makes that practical for Claude Cowork users.

cowork-research soft-depends on cowork-ai-os for identity (about-me/about-me.md, business-brain.md, connections.md). When cowork-ai-os v0.10.2+ is detected, the wizard's Phase 2 delegates connector recommendations to the live /browse-connectors skill — which fetches claude.com/connectors at runtime, so you always get the current catalog instead of a stale bundled snapshot. Older versions still work via the bundled fallback.

---

What it ships (4 skills)

Skill	Slash	What it does
Onboard wizard	/onboard-research	7-phase install (~12 min). Reads your business context, classifies into 1 of 6 business types, recommends top 3 connectors for your type, picks a default research destination, calibrates your preferences. State.md resumable.
Research brief	/research-brief	600-word executive brief + 5-bullet TLDR on any topic, competitor, account, methodology, or library. Cited. Plan-then-approve writes.
Web audit	/web-audit	CRO + competitive audit of any URL. Bulleted recommendations grouped by severity (block / fix / consider). Read-only on web.
Meeting recap	/meeting-recap	Fathom transcript → decisions + action items + open questions + notable quotes. Paste a Fathom URL or call ID and it does the rest.

That's the whole surface. Four commands. Run /onboard-research once and the wizard wires the rest.

---

How memory works (the 4 tiers)

Borrowed from cowork-obsidian's memory architecture. Same pattern, applied to research outputs instead of vault notes.

Tier	What	Where	Loaded when	~Token cost
1	Identity	about-me/ (cowork-ai-os)	Skills that need it	300
2	Research hot cache	reference/_research-hot.md	Start of any research session	500
3	Research index	reference/research/_index.md	When picking which past brief is relevant	1,000
4	Individual outputs	reference/research/briefs/<topic>.md (and audits/, meetings/)	Only when explicitly named or matched	600–1,200 each

A typical research query touches ~3,500 tokens, not 50,000. Past briefs become input context for new ones. Audits inform briefs. Meeting recaps surface in account research. The library compounds.

---

Per-business-type connector matrix

Phase 2 of the wizard reads your business-brain.md, classifies you into one of 6 business types, and recommends the top 3 connectors for that type. Here's the matrix: