I used to read research summaries that felt tidy and still left me unsure what I could actually rely on. The clean paragraphs mixed things the source had stated outright with conclusions the summary writer (or the model) had drawn, and the questions that remained unresolved simply disappeared. When I later tried to use the notes, I had to go back to the original document to untangle what was solid, what was interpreted, and what was still unknown.
The prompts below exist to stop that mixing. They force a three-way separation—facts, inferences, and open questions—so the output is harder to misread and easier to act on. They are short by design. Each one can be tested, tightened, or expanded without rewriting the entire instruction set.
Why the Separation Matters
Most professional research tasks are not about producing a polished essay. They are about extracting usable knowledge under time pressure. In that setting the most expensive errors are quiet ones:
Treating an inference as a stated fact
Missing a limitation that changes the scope of a claim
Forgetting that a key point was never resolved in the source
A summary that blends all three creates false confidence. A summary that keeps them visibly separate lets you decide, at a glance, what you can take to a meeting, what still needs checking, and what should remain an open item on your list.
The prompts that follow treat that separation as the primary job. Style, length, and tone are secondary.
Baseline Prompt (Start Here)
Summarize the following material.
Separate the output into three labeled sections:
Facts (directly stated in the source)
Inferences (reasonable conclusions that are not explicitly stated)
Open questions (important points the source does not resolve)
Rules:
Keep important exceptions or limitations with the relevant facts
Do not move a statement from Facts to Inferences (or the reverse) unless the source clearly supports the change
If a section would be empty, write “None identified” rather than filling it
This is the version I reach for first. It works on articles, reports, meeting transcripts, and long emails. The three labels make the status of every statement visible. The empty-section rule prevents the model from inventing content just to populate a heading.
Variation 1: Stricter Fact Definition
Use when previous outputs have been too generous about what counts as a fact.
Summarize the material using three sections: Facts, Inferences, Open Questions.
Treat a statement as a Fact only if it is explicitly asserted in the source.
Any conclusion that requires combining statements or reading between the lines belongs under Inferences.
Any material gap that affects the reliability or scope of the claims belongs under Open Questions.
Preserve exceptions with the facts they modify.
Variation 2: Source-Linked Facts
Use when you need to return to the original passage quickly.
In the Facts section, add a short location note after each fact (page, section heading, or paragraph number if available).
Keep Inferences and Open Questions unlabeled by location unless the source position is essential to understanding the gap.
Do not invent locations.
This version is slower to read but faster to verify. I use it when the stakes of a misread are high.
Variation 3: Decision-Focused Open Questions
Use when the research is meant to support a specific choice.
After the three standard sections, add a fourth section titled “Open Questions That Affect the Decision.”
List only the unresolved points that would change the choice if answered differently.
If no open question reaches that threshold, write “None identified.”
The extra filter keeps the open-questions list from becoming a dumping ground for every possible curiosity.
Variation 4: Multiple-Source Version
Use when you are summarizing several documents at once.
Summarize the following sources.
Produce three sections: Facts, Inferences, Open Questions.
For every Fact, note which source it comes from.
If sources conflict on a factual point, record the conflict under Facts and describe it briefly.
Keep Inferences clearly marked as interpretations, not as additional facts.
Conflicts are easy to smooth over in a blended summary. This version forces them into the open.
Variation 5: Length-Controlled With Safety Flag
Use when the output must stay short.
Produce a concise three-section summary (Facts / Inferences / Open Questions).
Hard limit: [set your limit, e.g., 250 words].
If an important exception or limitation must be cut to meet the limit, replace it with the note “[exception removed – see source]” instead of dropping it silently.
Visible omission is safer than invisible omission. I use this when space is genuinely constrained and I still need to know what was left out.
Variation 6: Diagnostic Self-Check
Use when you want the model to surface its own uncertainty.
After completing the three sections, add a short note titled “Separation Confidence.”
State which, if any, statements were difficult to classify and why.
Do not change the classifications—only flag the difficulty.
This is a diagnostic aid, not a replacement for human review. It often highlights the exact sentences I should double-check first.
How to Choose and Adapt
I keep the baseline and two variations in active use. The others stay in a reference note for the document types that need them. When I adapt a prompt, I change only one element at a time—usually the strictness of the fact definition or the addition of source location—and compare the new version against the baseline on the same material.
A practical way to start:
Run the baseline on a document you already know well.
Check whether any fact is actually an inference, and whether any important open question disappeared.
If the separation is too loose, try Variation 1 on the same document.
Keep the version that produces the cleanest, most honest division with the least editing.
That single comparison teaches more than collecting another set of templates.
What These Prompts Cannot Do
They cannot turn a poorly sourced document into a reliable one.
They cannot judge which open questions matter most for your specific decision—that judgment stays with you.
They cannot replace reading the source when the cost of error is high.
Every output is a draft. I read the Facts section against the original before I rely on it, and I treat the Inferences section as a set of hypotheses rather than conclusions. The prompts reduce the chance that the three categories will be quietly mixed; they do not eliminate the need for human review.
Why Shorter Prompts Help Here
I have tested longer research prompts that asked for tone, prioritization, recommendations, and confidence scores in one pass. The outputs looked more complete. They also blurred the boundaries between fact and inference more readily and required more cleanup. The short baseline above produces a cleaner separation and leaves the judgment about what matters where it belongs—with the person who has to use the notes.
When a repeated problem appears (for example, technical terms being simplified out of recognition), I add one targeted instruction and test it. Most additions do not survive the comparison. The few that do earn their place by solving a concrete failure mode.
Building Your Own Version
Copy the baseline and run it on two or three real research documents this week. After each run, note where the separation failed or where you had to move statements between sections. Those failure points are candidates for a single new instruction. Add the instruction, test it, and keep it only if the separation becomes more reliable without creating new problems.
The goal is not a perfect research assistant. The goal is a short, repeatable method that keeps facts, inferences, and open questions visibly distinct so you can decide what is solid, what is provisional, and what is still unknown.
Start with the smallest useful version—the three-section baseline that refuses to leave any section empty by invention. Test it on real material. Add complexity only when the documents themselves prove you need it. The notes stay a little less polished and a lot harder to misread. That trade-off has held up better than any longer, more ambitious research prompt I have used.
No notes on this sheet yet.