🟡 The citation machine
You get: architecture decision records with receipts. Pairs with: the colleague who asks “source?” in every review.
tabstack research "CRDTs vs operational transforms for a collaborative editor in 2026"
> /tmp/r.ndjson
jq -r 'select(.event=="complete") | .data.report' /tmp/r.ndjson > docs/adr/007-crdts.md
jq -r 'select(.event=="complete") | .data.metadata.citedPages[] | "- [(.title)]((.url))"'
/tmp/r.ndjson >> docs/adr/007-crdts.md What it does
research runs a multi-source investigation and returns a synthesized report with full citation metadata. The NDJSON stream carries everything — pipe to a file, then extract the report and citations separately with jq.
The result is an ADR that cites its sources, automatically.
Why save to a file first
Piping to a file lets you query the NDJSON multiple times without re-running the API call (~250 credits for research). Extract the report once, the citations once, anything else you need — all from the same stream.
The cost note
research costs roughly 25× more than extract. If you can name the sources in advance, extract each page directly and synthesize in-context — you’ll often get better specificity at a fraction of the price.
Use research when the hard part is finding the right sources, not reading them.
Extend the output
# Add a "further reading" section with abstracts
jq -r 'select(.event=="complete") | .data.metadata.citedPages[] |
"### (.title)\n(.url)\n"' /tmp/r.ndjson >> docs/adr/007-crdts.md One-liner version
When you don’t need to save the stream:
tabstack research "WebSockets vs SSE for real-time collaboration"
| tee /tmp/r.ndjson
| jq -r 'select(.event=="complete") | .data.report' tee saves to file and passes through to jq simultaneously.
Good questions for this recipe
- “What are the tradeoffs between X and Y for [specific use case] in [year]?”
- “What does current research say about [technical topic]?”
- “What have teams learned from adopting [technology] in production?”
Broad discovery questions where you don’t know which sources to read first. Once research surfaces them, extract the cited pages directly for the detail you need.