Pathwise

Prompting AI: From First Question to Reliable Results · Lesson 8 of 12 · 12 min

Make it check itself, and you check it

Ask for reasoning, assumptions and uncertainty so you can see where an answer might be wrong. Then verify what matters in a primary source, and never trust a citation you haven't opened.

LAYER ONE

Ask for the steps and the assumptions

For anything with logic or arithmetic, ask the model to work through it step by step before giving the answer. This tends to improve the result, because each step becomes text that the next step can build on. It also lets you see where things went wrong. Add: "List every assumption you made." Assumptions are where a sensible-looking answer quietly stops being about you.

The steps are something to inspect, not proof. An explanation can sound tidy and still contain a wrong step, so read it like a colleague's working, not like a certificate.

A monthly transport pass costs 900. A single trip costs 25. Is the pass worth it for me? Work step by step, and list every assumption you make.

Output

Assumptions:
- You commute 22 working days a month.
- You make 2 trips per day.
- No other trips.

Steps:
1. Trips per month: 22 x 2 = 44
2. Cost without pass: 44 x 25 = 1,100
3. Pass costs 900, so it saves 200.

Answer: Yes, the pass is worth it under these assumptions.

The arithmetic is right. But suppose you work from home two days a week: about 13 days, 26 trips, 650. The pass loses. Without the assumptions list you would have seen only "Yes".

Check yourself

In the example above, what did asking for assumptions actually give the user?

  1. A way to spot that "22 working days" doesn't match their life, before acting on the answer
  2. A guarantee that the final answer is correct
  3. A longer answer, which is always more reliable
  4. Proof that the model did the maths properly
Show the answer

A way to spot that "22 working days" doesn't match their life, before acting on the answer

Right. The model had to guess what it wasn't told. Listing the guesses puts them where you can see and correct them.

LAYER ONE, CONTINUED

Ask how sure it is

By default an assistant sounds equally sure about everything. Make it sort its own claims: "Mark each point as confident, fairly sure, or a guess", or "Which parts of this should I double-check?" The labels aren't perfectly calibrated, so a "confident" can still be wrong. But a "guess" label is a gift: it tells you exactly where to look first.

Also worth adding to factual questions: "If you don't know, say so instead of guessing." It gives the model a way out that isn't an invented fact.

Check yourself

Reza asks his assistant for the date a law came into force. He then asks, "Are you sure?" and it replies, "Yes, I'm certain." The date is now verified.

Show the answer

False

The second reply comes from the same model with the same knowledge. It's more text, not more evidence. A date like this is verified when Reza finds it in the law itself or on an official page.

Primary source

NOUN · RESEARCH

The original place a fact comes from, not someone's retelling of it: the text of the law, the statistics office's own table, the research paper itself, the company's own page, the contract you signed. An assistant's answer is never a primary source. At best it's a pointer towards one.

"A study showed that..." is a retelling. The study's own abstract and results are the primary source. Does it exist? Does it say that?

Layer two: your checking routine

  1. Pick the claims that matter

    You can't check everything. Check what you'll act on or pass on, starting with lesson 3's four risk zones: citations, exact numbers, niche facts, recent events.

  2. Ask where to look

    "Where could I verify this? Name the kind of source, not a link." A suggestion like "the national statistics office" is more dependable than a URL, which may be invented.

  3. Open the source yourself

    Search for it and get to the real page or document with your own hands.

  4. Find the exact sentence or number

    Not "the topic seems right". The actual figure, the actual clause.

  5. Can't find it? Don't use it

    An unverified claim goes out of your text, or stays in clearly marked as unverified.

Check yourself

Which of these count as verifying a claim?

  • Opening the cited article and finding the sentence
  • Asking the same assistant the question again
  • Checking the figure on the statistics office's website
  • Noticing that the link looks official
  • Reading the clause in the actual contract
  • Seeing that the answer includes a reference number
Show the answer

Real verification: Opening the cited article and finding the sentence, Checking the figure on the statistics office's website, Reading the clause in the actual contract

Only feels like it: Asking the same assistant the question again, Noticing that the link looks official, Seeing that the answer includes a reference number

The citation rule

  • Never pass on a citation you haven't opened. Not in an essay, a report, a post or a message to your boss.
  • Check that the source exists, and that it says what the answer claims. Real sources get misquoted too.
  • Even when a tool searches the web and shows links, click through. The summary beside a link can still be wrong.
  • If checking would take longer than the claim is worth, drop the claim.

Check yourself

Match each worry to the prompt that addresses it

Show the answer
  • "It may have guessed things about my situation" → "List every assumption you made."
  • "I can't tell the solid parts from the shaky ones" → "Mark each point: confident, fairly sure, or guess."
  • "This calculation might be wrong" → "Solve it a second way and compare the results."
  • "It might invent a fact to fill a gap" → "If you don't know, say so instead of guessing."

Check yourself

Negar's thesis chapter needs sources. An assistant gives her five references, each with authors, journal, year and pages. She has one evening. What should she do?

  1. Ask a second assistant whether the references look real
  2. Use all five, since the details are too specific to be invented
  3. Use them, but ask the assistant to format them properly first
  4. Search for each one, keep only those she can open, and read the relevant part before citing it
Show the answer

Search for each one, keep only those she can open, and read the relevant part before citing it

Yes. References are the single riskiest output. Any she can't find is out, and any she can find must actually say what she claims it says.

Lesson recap

  • Ask for step-by-step working and a list of assumptions, so errors and hidden guesses become visible.
  • Ask which parts it's unsure of, and allow "I don't know". These are useful signals, not guarantees.
  • "Are you sure?" verifies nothing. A second method or an outside source does.
  • Check the claims that matter in a primary source, and never pass on a citation you haven't opened.

Keep it, don't just read it

Pathwise brings each idea back just before you'd forget it, with a quick question. Free on Android and on the web, in English and Persian.

Cafe Bazaar Myket Open the web app

All lessons in this course

  1. It predicts the next word
  2. Context is everything
  3. Why it makes things up
  4. The four parts: role, context, task, format
  5. Show, don't just tell
  6. Constraints and iteration
  7. Break big tasks into steps
  8. Make it check itself, and you check it
  9. Working with long text
  10. Writing and editing with AI
  11. Learning anything faster
  12. Your prompt library and staying safe