Claude's long context is genuinely useful for whole-document work: contract review, specification analysis, synthesising a stack of reports. It is also easy to use badly.
Put a map at the top
Begin with a short table of contents or a one-paragraph description of what the document contains and what you want from it. This measurably improves how well later sections are used.
Ask for one thing at a time
A single prompt requesting a summary, a risk list, a rewrite and a set of recommendations produces a shallow answer to each. Separate passes produce far better output, and the intermediate results are easier to check.
Work with section markers
Explicit markers between sections let you refer to "section 4" in follow-up questions instead of quoting text. It keeps turns short and keeps the model anchored.
Verify the specifics
Long-context recall degrades in the middle of very large inputs. For anything consequential, ask for the supporting quote alongside each conclusion and check it against the source.
Comments (0)
Log in to join the discussion
Log InNo comments yet