Get a real synthesis out of a document too long to read yourself

Why this works

Claude's context window is long enough to hold an entire contract, transcript, or research paper at once, but a bare summarize this invites the model to weight the parts it read most recently or most confidently, producing a synthesis that skews toward the document's opening sections — a documented recency and salience bias in long-context summarization. Requiring the answer up front and the evidence with location citations second forces two separate passes: commit to a claim, then justify it against the actual text, which is harder to fake than a single pass that blends recall and inference into one paragraph. The instruction to surface contradictions by name rather than silently reconciling them directly targets long-document Claude's tendency to smooth over an internal inconsistency into one coherent-sounding story, which is exactly the failure mode that matters most in a contract or policy document where two clauses genuinely conflicting is the whole point of asking. Explicitly forbidding an answer from general knowledge closes the other common failure: for a well-known contract type or topic, the model has prior training knowledge about how such documents usually work, and without this constraint it can blend that prior with the actual document instead of grounding purely in what was pasted.