No. 011
Stop AI making things upin your business work.
Chin Gregory presents · a guide on ai tools · runtime 10 min
Free guide · No. 011 · AI Tools

Stop AI making things up in your business work.

Four prompts that make Claude or ChatGPT cite its sources, say when it doesn’t know, check its own work and keep facts apart from guesses. Written around the work you’re doing.

Takes
10 minutes
Needs
Claude or ChatGPT
Costs
Nothing. Works on the free plans

Get answers you can put in front of a client without wondering which number was invented. Four prompts make Claude or ChatGPT show where every fact came from, admit what it doesn’t know, and mark its guesses. Fill in the form and they’re written around your work.

What you’ll need.

Get these set up first. Everything below uses them.

Some of these are affiliate links. If you sign up through one, I may earn a commission, at no extra cost to you. See every tool I use.

01

Why AI makes things up

One paragraph. Knowing why makes the fix obvious.

Claude and ChatGPT write by predicting the most likely next words, and the way they’re trained rewards a confident guess more than “I don’t know” (OpenAI’s own research says exactly this). So when they’re missing a number, a quote or a source, they can fill the gap with something that sounds right. The fix is to give them permission not to know, give them your sources, and make them show their working.

02

Tell this page what you’re working on

Your answers write all four prompts. They stay in your browser and nobody else sees them.

A proposal, a client report, a course module, an email, a blog post.
The things that would embarrass you or cost you money if they were wrong.
Files you’ll attach, or websites you trust.
03

Prompt 1: set the ground rules

Rules at the start of a chat shape every answer after it. This one gives AI permission to say “I don’t know”.

  1. Start a new chat in Claude or ChatGPT.
  2. Attach your sources if you have them: your price list, notes, a report, a transcript.
  3. Paste this, send it, and wait for “Got it”. Then ask for what you need.
Your ground rules
Before we start, here are the rules for this chat.

I'm working on: [WHAT YOURE WORKING ON]
These things must be accurate: [WHAT MUST BE ACCURATE]

1. If you don't know something, say "I don't know". A gap is fine. A made-up answer isn't.
2. Never invent a number, date, name, quote, price, statistic, study or link. If one belongs somewhere and you don't have it, write [NEED: what's missing] and keep going.
3. Use my sources first. If you use anything else, say where it came from.
4. Mark anything you're unsure of with (unsure) right after it.
5. If my question is unclear or you're missing something important, ask me before you write.

Reply "Got it" and wait for my first request.
Gaps now show up as [NEED] instead of a made-up number.
04

Prompt 2: make it cite its sources

A fact with a link is a fact you can check in ten seconds. A fact without one is a guess until you prove it.

When you need facts from the web

Both apps can search the web and link their sources. In Claude, it searches when it helps; if you see a Web search option under the + button, switch it on. In ChatGPT, it searches on its own too, or type / and choose Search.

Cite web sources
Search the web to answer this: [your question]

Context: I need this for [WHAT YOURE WORKING ON].

- Use official or original sources where you can: the company's own site, the government page, the original study.
- After every fact, put the source in brackets with its link.
- Tell me the date each source was published or last updated.
- If sources disagree, show me both and say which you'd trust and why.
- If you can't find a source for something, say "No source found" instead of answering from memory.

When the facts are in your own documents

Attach the files first. Pulling exact quotes before answering keeps the answer tied to what your document really says.

Cite your documents
Use only the documents I've attached. Don't use anything you already know.

Step 1: Find the word-for-word quotes that answer my question. Number them. If you can't find any, say "No relevant quotes found" and stop.
Step 2: Answer using only those quotes. After each point, add the quote number in brackets, like [2].

My question: [your question]
Click the links. Search results and citations can still be wrong or out of date. Open the source and check it says what the answer says it does.
05

Prompt 3: make it check its own work

Checking every claim against a source is one of the ways Anthropic recommends to cut down made-up answers.

Send this straight after any answer you plan to use. Same chat.

Check your own work
Now check your last answer like a strict editor.

1. List every factual claim in it: every number, name, date, price, quote and statement of fact. Pay most attention to: [WHAT MUST BE ACCURATE].
2. For each one, tell me where it came from: a source I gave you (quote it), a web page (link it), or your own memory.
3. Mark each one: CHECKED, UNSURE or MADE UP.
4. Rewrite the answer. Keep the CHECKED claims. Remove the MADE UP ones and put [REMOVED] where they were. Put (unsure) after the UNSURE ones.
5. Finish with a list of everything I should check myself before I use this.
You get a cleaned-up answer and a short list of what to check yourself.
06

Prompt 4: keep facts apart from guesses

Use this for planning and advice, where a guess dressed up as a fact does the most damage.

Separate facts from guesses
Split your answer into three parts.

FACTS: things you can back up with a source I gave you or a link. Put the source after each one.
GUESSES: things you think are likely but can't prove. Say why you think each one.
QUESTIONS: what you'd need from me to turn a guess into a fact.

Keep the three apart. Don't let a guess sit in the facts.
Guesses are useful. An educated guess is worth having. You just need to know it’s a guess before you act on it.
07

Do the last check yourself

The prompts cut the made-up stuff down a lot. The final check on anything that matters is yours.

  1. Open every link. Make sure the page says what the answer says it does.
  2. Search the source for every quote. It should be there word for word.
  3. Check every number against where it came from.
  4. Fill or cut every [NEED] and [REMOVED] gap.
  5. For anything big, ask the same question in a new chat. If the two answers disagree, at least one of them is wrong.
That’s it. You can stand behind every fact in it.
08

Take your downloads

Tick what you want. Each file is written from your answers in step 2.

Your tools, one more time.

If something goes wrong.

It gave me a link that doesn’t work

Reply: "That link doesn’t open. Search the web again and give me a working link for that fact, or tell me you can’t find one." Then remove the fact if it can’t.

It still made up a number

Paste Prompt 1 again and say: "You broke rule 2. Replace every number you can’t source with [NEED]." Long chats drift, so a new chat with the ground rules at the top helps too.

It says “I don’t know” to everything now

That means it needs more to go on. Attach your sources, or ask it to search the web. Then ask what it needs: "What would you need from me to answer this?"

The quotes from my document aren’t really in it

Check the file uploaded properly and isn’t a scanned image. Then use the "Cite your documents" prompt, which makes it pull exact quotes before it answers.

It agrees with me every time I push back

Ask it to check the source instead of your opinion: "Don’t agree with me to be polite. Check the source and tell me who’s right."

What’s next

Now use AI you can trust to sell.

Your AI now shows its sources and owns up when it doesn’t know. In my free live workshop I show you how I use Claude to sell my digital products and bring in customers every day.

Format
Free live workshop