Corpus summarisation — many docs to one report

TL;DR. Take 20–200 documents on a single topic and produce a structured synthesis — themes, evidence, gaps, sources. Weeks down to hours for due diligence; days down to hours for strategy. The 1M context window earns its keep here, for corpora over roughly 150 pages.

Job to be done#

Take a corpus of 20–200 documents and produce a structured Word synthesis with cited evidence and explicit gaps.

Who runs it#

Strategy team, legal reviewing case files, M&A team in due diligence, research lead doing literature reviews.

Inputs (inbox/)#

• The corpus in /inbox/corpus-[topic]/
• A research question or hypothesis
• A target output structure (or "use the standard")

Outputs (output/)#

• synthesis-[topic].docx — 6–10 page Word doc
• evidence-table.xlsx — claims with sources
• gaps-and-questions.md — what the corpus did not answer

Prompt seed#

Read every file in /inbox/corpus-[topic]/.
Goal: answer the question in /inbox/research-question.md.
Produce /output/synthesis-[topic].docx with sections:
- Executive summary (250 words)
- Themes (3–5, each 1–2 paragraphs)
- Evidence per theme (linked to sources)
- Confidence and gaps
- Recommended next questions
Generate /output/evidence-table.xlsx with one row per claim
(theme, claim, source filename, page, confidence).
Cite every claim. Mark inferred claims as such.

Quality bar#

• Every claim cited or marked inferred.
• Gaps section is meaningful, not vestigial.
• Common trip-up: confident synthesis of a thin corpus. Push back when 5 documents become 5 themes — the right output for a thin corpus is "this corpus does not support a synthesis."

Time saved (typical)#

Weeks down to hours for due-diligence corpora; days down to hours for strategy literature reviews.

Upgrade path#

• Connector to a document warehouse (SharePoint, Box) for the source pull.
• A corpus-synthesis skill with the standard structure baked in.

Tinkso's take#

This is the workflow where Cowork's 1M context window earns its keep. For corpora over roughly 150 pages, run with the full window. For smaller corpora, chunk and let Cowork synthesise across summaries to save quota — chunking is faster, cheaper, and often more accurate when there's no genuine cross-document reasoning required.