Ask Claude to write something from scratch and you'll get competent, generic output. Give it your messy first draft and ask it to cut, tighten, and clarify - the results are noticeably sharper. That gap, observed after eight months of daily use for writing and research, points to something real about how large language models work: they perform better when reacting than when inventing.
The blank-page problem isn't unique to Claude. But the difference between "write me a piece about X" and "here's my rough take on X, make it better" is wider than most users expect. When editing, the model has something concrete to push against. Generation from nothing pulls it toward a safe, averaged version of what that type of content usually looks like - the statistical center of a million similar pieces.
The Long Context Trap
The second finding cuts against how many people use Claude: loading up the context window with everything you have doesn't improve the output. Dumping 40 pages into a conversation and asking a question gets you a response trying to account for too much at once. The model doesn't automatically weight the most relevant passages higher - more material can dilute focus rather than sharpen it.
The practical fix is deliberate selection. Paste the specific passage relevant to your question. If you're working with a long document, query specific sections rather than feeding the whole thing upfront. Claude's context window can hold around 200,000 tokens (roughly a 500-page book's worth of text), but that's a ceiling, not a target.
Both observations point toward the same underlying principle: Claude works better with constraints. A defined scope, a specific excerpt, a rough draft to react to - these reduce the solution space and produce tighter output than open-ended prompts do.