ChatGPTSummarizationLevel 2: Context

How Do You Get a Summary That Helps You Decide?

Name the decision in your prompt, and ask only for what bears on it: the deciding facts, what argues against your current lean, a recommendation, and its main risk. "Summarize this" gives you a shorter document that you still have to interpret.

Prompt to copy
I need to decide: [the decision, with the options].
My current lean: [which way, or "none"].
What matters most to me: [for example, cost over three years, risk of lock-in].

Use only the attached document. Give me:

1. The 3 to 5 facts that bear on this decision, each with the section or page where it appears.
2. The strongest point against my current lean.
3. Anything the document doesn't say that I'd need to know before deciding.
4. Your recommendation in one sentence, and its main risk.

Leave out everything that doesn't affect the decision. If a point is your inference rather than something the document states, label it "Inference".

[attach the document or paste it here]
  • The deciding factsOnly what changes the outcome
  • The case againstWhat argues with your current lean
  • A recommendationOne clear call
  • Its main riskWhat could make it wrong
What to ask for once you've named the decision. A rule of thumb, based on OpenAI's prompting guide.

Try it now

  1. 1Attach a document you need to act on and paste the prompt above with your decision filled in.
  2. 2Check two of the facts against the pages it names.
  3. 3Send this follow-up: "What assumptions are you making? Which one would change your recommendation?"
  4. 4It worked if you could make the call, or know exactly what to find out next, without reading the whole document.

Use it when

Vendor proposals, contracts, research reports and long email threads when you have to act and can name the choice: renew or switch, sign or push back, hire or wait.

Skip it when

When you need a faithful neutral abstract, such as a literature review, minutes or a record for others. A decision frame would bend it. Also avoid using this as the only input on a decision with legal or medical stakes. Use it to prepare for a professional's review.

Why it works

A neutral summary treats every section as equally worth keeping. Naming the decision gives ChatGPT a way to choose, since a fact either changes the outcome or it doesn't. OpenAI's prompting guide says to start with the result you need and to state how you'll use the output, so it can set length and detail. Its suggested prompt for comparing options asks for the differences, a recommendation and the tradeoff.

Asking for counter-arguments matters because your framing pushes the answer. If you say you're leaning toward switching vendors, the model has a reason to lean that way too. OpenAI's follow-up "What assumptions are you making?" helps here. Its guidance also says to verify important numbers, names, dates, quotes and claims, and to have it state what it couldn't access.

Example

Say you have a 20-page renewal proposal from your software vendor, and you're leaning toward switching. You ask about three-year cost and lock-in. A useful answer gives the renewal price with its page, the auto-renewal notice period, what the proposal says about exporting your data, and what it leaves out, such as a price cap after year one. It may also point out that the migration cost isn't in the document at all, so you'd need to find that yourself.

Going further

For a decision built on conflicting sources, try a stronger setting. OpenAI's model-selection page suggests GPT-6.1 Sol at Extra high for decisions built from conflicting evidence. In Work, you can ask for the result as a memo, and OpenAI lists researching a decision with a memo among its examples. Ask for two versions from the same read: one line for a busy approver and half a page for the record. Then send only the one that suits each reader.

Common mistakes

  • Asking for "a summary" when you have a decision, then doing the interpreting yourself.
  • Stating your lean and not asking for the case against it.
  • Skipping what the document doesn't say. The missing clause can matter more than anything it includes.
  • Trusting a recommendation without checking two or three of its facts against the source.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course