You're using Claude to review a 1,200-line consulting proposal before it goes to a client. In a single-pass review, the agent's comments are shallow on individual sections and miss inconsistencies between sections — recommending a 6-month timeline on page 8 and a 4-month one on page 22 without flagging the contradiction. Sometimes the agent contradicts its own page-1 feedback by page 30.
What's the right architectural fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Multi-pass review separates local and global concerns. A per-section pass produces deeper local feedback because attention isn't diluted across 1,200 lines. A separate cross-section integration pass exists specifically to catch the inconsistent timelines, contradictory recommendations, and conflicting data that single-pass review systematically misses. Each pass has a coherent role; feedback stops contradicting itself.
A larger context window lets the model see more, but doesn't solve the lost-in-the-middle effect or attention dilution. Single-pass review of long documents produces shallow feedback regardless of context size.
The same session that produced shallow page-1 feedback won't catch its own shallow page-30 feedback by reviewing itself. Self-review confirms more often than it catches.
Sometimes the proposal really does need to be 1,200 lines. The architectural fix is multi-pass review; making the user shorten their document to fit your review approach is the wrong direction.