AI Guides › Claude Mastery
Why Long Documents Need A Different Prompting Style Than Short Questions
By Nigel Guy · 2 min read
People develop a prompting style on short questions — a quick line, an
answer, done — and then use exactly the same style once they upload a
ninety-page report. "What do you think?" works fine for a quick factual
question. Pointed at a long document, it produces something technically
responsive and practically useless: a generic overview that could have been
written from the title page alone.
The rule: a short question needs a sharp instruction; a long document
needs a scoped one — tell it where to look and what job to do with what it
finds, or it will guess at both.
The mechanism: the Scope/Job/Format split
Before sending a prompt with a long document attached, fill in three things
rather than one:
- Scope — which part of the document actually matters for this
question? "The whole thing" is a valid answer sometimes, but say so
deliberately rather than by default. Naming a section, a date range, or
a specific concern ("just the payment terms," "anything about
termination") does more work than any amount of politeness in the
phrasing.
- Job — what should happen to that material: summarise, compare
against something else, find inconsistencies, extract specific data,
argue a position? A long document can support any of these, but it needs
telling which one, because "what do you think" leaves the job undefined
and gets a job-neutral answer back.
- Format — how do you actually want the answer shaped? A table of
clauses and page references reads completely differently from a prose
paragraph, and for anything you'll refer back to later, the structured
version is usually the one you'll actually use.
Why this matters more as length grows
With a short question, an underspecified prompt mostly costs you a
follow-up message. With a long document, an underspecified prompt costs you
a full pass over material that may run to tens of thousands of words,
producing an answer shaped by whatever the model guessed you meant — and
you may not notice the guess was wrong until several exchanges later.
What to skip
Skip re-explaining the whole document's context in your prompt if you've
already uploaded it — the file carries that; your prompt only needs to carry
the scope, job and format. And skip breaking a genuinely unified document
into disconnected fragments to "make it easier" — that usually removes the
connective material (cross-references, running totals, an argument that
builds across sections) that made the long-document approach worth doing in
the first place.
Guardrails
- Very long documents can still exceed what fits in a single conversation
depending on length and format — if something seems to have been skipped,
ask directly rather than assuming full coverage.
- Scoping a request narrows what gets read, which is the point, but it also
means anything outside that scope wasn't checked — don't treat a scoped
answer as a full-document review unless you asked for one.
- For anything you'll rely on afterwards, ask for the answer to reference
specific sections or page numbers, so you can verify against the source
rather than trust the summary at face value.
All 751 AI guides · JulieMango plans from £17/mo