Future AI Pulse
Illustrative reading-workspace photograph by Yen Vu / Unsplash, used under the Unsplash License.
Quick answer: Ask an AI assistant to find short passages that answer your specific question before it writes a summary. Give every passage a location, then check the answer against those passages. This makes unsupported conclusions easier to spot.
Why use two passes?
Anthropic’s September 2023 long-context case study explored extracting relevant quotations before answering questions. It reported benefits in its particular experiments with older Claude models. That historical result is useful motivation; it does not prove this prompt improves every current model or document.
The practical advantage of the workflow below is inspectability. A polished summary can hide the point where an assistant jumps from a source statement to its own assumption. An evidence table gives you something concrete to check before accepting that jump.
Pass one: make an evidence table
Supply the document text or a readable attachment and name the question. If you have several documents, assign each a stable label such as “Release notes A” or “Manual B.” Preserve page numbers or section headings when available.
Question: [the exact question I need answered] Documents: [labelled source text or readable attachments] Find only passages relevant to the question. Return a table: document label | page/section | short excerpt | what it supports. Keep excerpts brief. Do not invent page numbers or quote unavailable text. If the answer is absent, say which evidence is missing. Do not write the final answer yet.
Pass two: answer within the evidence
Using the verified evidence table, answer my question in 150 words. Attach a document label and location to each factual claim. Separate explicit statements from your interpretation. List unresolved conflicts between documents. Do not fill gaps using general knowledge unless I ask for it.
For example, a release note saying a feature is “available to selected accounts” supports a limited rollout. It does not support “everyone has access.” If another document lists a broader rollout, compare their dates and scope instead of silently combining them.
A five-minute verification routine
Open each cited location and confirm that the excerpt is real, relevant, and not stripped of a qualifying sentence. Check at least one question whose answer you already know and one question the document cannot answer. The latter tests whether the assistant admits missing evidence.
If a scanned PDF produces broken text, fix the extraction or narrow the input before relying on the table. This is an original prompt template and suggested workflow, not a published benchmark or a substitute for reading consequential material yourself.
Sources
Prepared with AI assistance from the linked sources. Prompt templates and analysis are original editorial suggestions; no hands-on product testing is claimed.
Related reading
For another way to evaluate AI-assisted work, read our Scion coverage and infrastructure evaluation prompt.