The LIBRAiRY · Reference Shelf · unlocked in Episode 3

How to Check AI's Work

How to check an Ai answer without redoing all the work yourself — the move Elle Woods would make. Put the answer on the stand; don't let a claim speak without a receipt.

Ai can sound right, and sounding right is not the same as being right — the two come out in the same handwriting. This is the short, reusable method for telling them apart before your name goes on it. Steal it, keep it, run it on anything.

First

Sort what you're holding.

Before you check a single thing, name what it is. Every line an answer hands you is one of three things.

Draft

Wording, structure, a summary, a brainstorm. Nobody gets embarrassed by a rough phrasing.

Low stakes — can be fast.

Claim

A name, a date, a number, a quote, a source — anything legal, HR, privacy, or money.

This is what gets you — check it.

Receipt

The thing you can open, name, or point to if someone asks where a claim came from.

What every claim needs before it speaks.

A draft is an outfit. A claim is an alibi. Dress accordingly.

Prompt Like Elle

The three moves.

The serious guidance — OpenAI, Anthropic, Google, Stanford — all points the same way. In normal-human terms, it comes down to three moves.

Give it the source.

Don't ask the tool what it remembers; hand it the material. Paste the document, upload the file, link the official page — or turn on search mode when you need something current. Then set the boundary. Elle doesn't argue from memory; she walks in with the file.

Say: "Answer only from the document I gave you. If it isn't in there, say so."

Let it say "I don't know."

Add the one sentence most people never think to add. That single line is what turns a confident guess back into an honest blank.

Say: "If you aren't sure, say so. Don't guess to be helpful. Mark anything you inferred."

Make it show the line.

A claim without a source it can point to doesn't get to speak. If it can't quote the line, treat it like it showed up at Spring Fling with no student ID.

Say: "For every factual claim, quote the exact sentence you relied on."

One rule over all three: no invented receipts.

Say it out loud in the prompt — "Do not invent sources, links, dates, quotes, numbers, or policies. If a receipt is missing, mark it [needs receipt]." "Sources attached" is not the same as "sources checked." Sometimes it's still Chutney, handing you a folder in her own handwriting.

Copy · paste · done

The Prompt Like Elle template.

Drop your source in the quotes, swap in your task, send.

Your prompt — copy it, paste it into your AI tool

Here is the source material:
"""
[paste your document, notes, or page here]
"""

Using ONLY what's above:
1. [Your task — e.g. "turn these notes into a client update."]
2. If something isn't in the source, say "not stated." Do not fill the gap.
3. For every factual claim, quote the exact line you got it from.
4. Do not invent sources, links, dates, quotes, or numbers.
   Mark any missing receipt as [needs receipt].

When a claim still needs checking

Check patterns.

You won't verify everything — just the claims. Point each one at the question it can't dodge.

If it says a policy changed — what date proves it?
If it says a quote came from a person — where's the quote?
If it says a number moved — from what, over what timeframe, according to whom?
If it says "best practice" — whose? For whom? Under what constraints? In what jurisdiction?

Trust a citation only when you've clicked it, it exists, and it actually says what the answer says it says. Otherwise: Google it, or ask the human who owns the decision. Make the second check independent — don't let the same witness validate itself.

The Chutney test.

Don't ask "does this sound smart?" — Chutney sounded smart three times. Ask: what's the one detail that can't survive contact with the timeline? Find it, and the answer either holds up or falls apart in your hands — before it ever leaves your laptop.