How to Scope AI Context to Selected Sources

A prompt such as “research this topic” hides the most important editorial decision: which material is allowed to influence the answer. Write a one-sentence research question, then list the sources that may answer it. If you are comparing an explainer video with a policy document, name those two items and exclude background notes, old drafts, and unrelated board material from the first pass.

Give the boundary a purpose as well as a list. “Use Source A for the speaker's explanation and Source B for the formal definition” is more useful than “use both sources.” Google’s Gemini Notebook help makes the same practical distinction: when several sources are selected, naming them in a specific question helps narrow retrieval. Treat that as a workflow technique, not proof that any tool will obey perfectly.

Use one small record for each source: source ID, author or publisher, title, date, format, rights or access note, and the question it can answer. The ID can be plain—PDF-1 or VIDEO-1—but it must remain stable in prompts, notes, and the final fact check. A filename such as final-v7.pdf is not a useful identity once several versions appear.

Add a condition field for what the AI actually receives. A PDF may contain selectable text, page images, or both. A video workflow may receive a transcript rather than the visual track. Record that distinction before asking questions, because it changes which claims can be supported. The inventory prevents a transcript-derived answer from being described later as full video analysis.

Run a source-specific pass for each item. Ask for the source's answer to the research question, the strongest passage or segment, qualifications, and anything the source does not establish. Do not ask for polished prose yet. The output you want is a compact evidence note that can be checked against the original.

Use the same question structure across sources so differences remain visible. If Source A uses a term differently from Source B, preserve both definitions rather than asking the model to reconcile them immediately. Early reconciliation often erases the exact disagreement that makes a comparison valuable.

Open the original at the supplied page or timestamp. Check whether the wording, speaker, and nearby context match the note. For a PDF, inspect tables, footnotes, captions, and page headers when they affect meaning. For a video transcript, replay the segment and look for an on-screen chart, demonstration, or edit that the text does not capture.

Classify each candidate as supported, qualified, conflicting, unsupported, or not checked. “Qualified” is useful when the source supports the core statement only under stated conditions. “Unsupported” means the current source pack does not establish it; it does not necessarily mean the idea is false. This vocabulary keeps absence of evidence separate from contradiction.

Now ask the cross-source question using only the accepted and qualified notes. Require the synthesis to keep source IDs and locators attached to the statements they support. If two sources disagree, the output should state the disagreement and its likely cause—different scope, date, definition, or evidence—without inventing a winner.

Draft from the synthesis, then run a source-debt pass. Highlight every sentence that contains a checkable external claim. Each should point to an accepted evidence note, be rewritten as analysis, or be removed. The method cannot eliminate model error, but it makes unsupported transitions easier to see before publication.

Place this block at the top of the board or research note: Question; intended output; included source IDs; excluded context; required locators; claim-status vocabulary; and final reviewer. Reuse the block in the drafting prompt so the evidence contract stays visible when the work moves from extraction to writing.

When a new source becomes necessary, add it deliberately. Record why it entered, which gap it fills, and whether it changes any accepted claim. That small change log is more trustworthy than quietly expanding context until the answer sounds complete.