Research and synthesis at scale with Claude Cowork

TL;DR. Cowork compresses 4 hours of "open 12 tabs and read PDFs" into about 40 minutes of structured synthesis. It is genuinely good at corpus work, web + local mixes, and cross-document Q&A. The catch is calibration: the first three research outputs need a human checking citations, or the team learns to trust them faster than they should.

Three patterns Cowork is genuinely good at#

Document corpus synthesis. Drop 30 PDFs into a folder, ask for one structured report. Cowork reads them all, identifies recurring themes, surfaces conflicts, and writes a brief with source citations. This is the pattern that produces the dramatic time-saved numbers in our engagements.

Web + local mix. Cowork searches the web, pulls public sources, and joins them with internal documents in the same run. A competitive teardown becomes "search for the three competitors, pull their public pricing pages, cross-reference our internal sales notes, write the comparison."

Cross-document Q&A. "Find every contract in ~/contracts/ that mentions a force-majeure clause about pandemics." Cowork scans the corpus, extracts the relevant passages, and produces a list with citations. This used to be a paralegal week; now it is a coffee.

The corpus synthesis prompt pattern#

Read every PDF in ~/research/q2-competitors.
Produce ~/output/q2-competitive-brief.docx with this structure:
1. Executive summary (200 words)
2. Themes — 3 to 5, each with 2 supporting quotes and source filenames
3. Evidence table — Excel sidecar at ~/output/q2-evidence.xlsx
   (one row per claim: claim, source_filename, page, my_comment)
4. Gaps and open questions — what you could not verify or had to infer
5. Sources — every PDF you read, in order

Voice: British English, no marketing language.
Plan first; show me the theme list before drafting.

The structure is the work. Without it Cowork will produce a competent but generic synthesis that nobody wants to share. With it, the output is a brief your team can hand to the executive sponsor on the way to the meeting.

Web research limits to know#

• Cowork's web search is built-in. It does not log into anything; it reads what is publicly readable.
• Paywalled content is not bypassable. Cowork respects the wall. If you need paywalled sources, get them yourself and put them in the workspace.
• Forum and social content is reachable, but quality varies wildly. Always ask Cowork to disclose the source domain in the citation, so you can weight credibility yourself.

If your research depends on logged-in access to a system, that is what MCP connectors are for — see the next section.

The 1M context window in practice#

The big context window matters less for research than the marketing suggests, and more in narrow situations.

• Useful when the work is naturally one corpus — a single contract review, a single year of monthly reports, a single literature review. The cross-document reasoning is real and worth the quota cost.
• Less useful when the work is many small chunks. Chunking with sub-agents is faster, cheaper, and often more accurate.

Tinkso heuristic: if the corpus would print to 200+ pages, use the long context. Otherwise chunk and aggregate. The crossover is roughly where chunking overhead exceeds the loss of cross-document attention.

Citations and trust#

Every research artifact should include source citations. Cowork will add them on request — and will not add them by default if you forget to ask. Make it a CLAUDE.md rule.

The Tinkso convention: every research output ends with a Confidence and gaps section flagging anything Cowork inferred, anything it could not verify, and anything where two sources conflicted. The structure looks like:

1. Executive summary
2. Themes
3. Evidence table
4. Confidence and gaps
5. Sources

The fourth section is what separates a research output your team will trust from one they will quietly re-check.

When to combine with an MCP connector#

Connectors unlock research over content that is not in the workspace folder.

• Internal-only research — connect Cowork to your knowledge base via MCP: Notion, Confluence, SharePoint. The corpus is the company's own brain.
• Customer research — connect to your CRM (Salesforce, HubSpot) for account-level synthesis ("everything we know about Acme, plus a public-web check on what they have done since the last meeting").
• Hybrid research — local PDFs + internal knowledge base + web. The most powerful pattern, and the most fiddly to set up.

See MCP connectors for the build vs buy decision.

Tinkso's take#

Research workflows are where Cowork's hours-saved numbers look most absurd in client engagements — four-hour competitive teardowns into forty minutes, repeatedly. The catch is calibration: the first three reports need a human checking citations cold, or the team trains itself to trust outputs faster than the model deserves.

We treat the first ninety days of any research workflow as calibration mode. Spot-check every fifth output, write down the misses, edit the prompt. After ninety days the calibration is internalised and review can shrink. Skip the calibration phase and the team will eventually publish something embarrassing.

Try this#

Take three competitor PDFs. Run the corpus synthesis prompt above. Read the output. Then read the source PDFs cold. Note where Cowork over-claimed and where it missed nuance. Edit the Confidence and gaps instructions in your prompt to address the specific patterns you saw. That is how a generic prompt becomes your team's prompt.