The AI prompt that wins a yes, and one that cites sources

The Take

AI wants to look done.

Ask it for anything & it hands you something finished. A tidy report, a confident script, a plan that’s ready to run.

It looks like real work got done.

Some of it is. Some of it’s fiction. You can’t tell by looking.

Every prompt this week makes it deliver facts. Show your work, cite the source, ask me before you run off & build.

Turns out we’re all back in grade school, checking work.

This time we’re the teacher.

Seen This Week

The prompt that wins a yes without pushing

It’s Friday. You want out by 3 & the only person who can cover the desk is Sam.

So you do the lazy thing. You open the chat & type “Come up with something good to convince Sam to cover for me.”

It writes, “3 reasons Sam should say yes.” A closer so sure of itself you almost laugh.

You’d never say a word of it out loud. It sounds like a tele-marketer selling you life insurance.

And we all know people typically agree more when the whole thing feels like their idea.

That’s why this prompt doesn’t hand you a pitch.

It builds the ask around a question Sam answers, names the objection before they can & gives them 2 ways to say yes that work for everyone.

I need to get [person, like my manager or my partner] to [do the thing]. They haven't said yes yet.

First, ask me 2 things, then wait for my answer:
1. How big is the ask, and have we disagreed about it before?
2. Who is this person to me, a partner, a coworker, a friend? The same words land differently depending on that.

Then build the approach on how people actually agree to things:
- Open with a calibrated question, one that starts with "how" or "what" and makes them help solve my problem ("how am I supposed to get this done by Friday and still cover the desk?"). Never one they can answer with a flat "no."
- Name their likely objection before they raise it, so it loses its charge.
- Give them 2 options that both work for me instead of one demand. People move faster when the choice feels like theirs.

Give me 3 things I could actually say, shortest to longest, matched to who this person is to me. Tell me which to open with. Then a fallback if they still balk, the smallest yes that keeps things moving, not the whole ask at once.

Why it works: It asks 2 questions before writing anything, so the ask fits the actual person instead of a stranger. The calibrated question hands them your problem to solve instead of something to push back on. Naming the objection first takes the air out of it.

One thing: That same move backfires if you guess the objection wrong. Tell Sam “you probably think this is too much to ask” & you’ve just handed them a reason they hadn’t thought of.

Simple fix: Tell it to open with the question & hold the objection line until you’ve heard their real one.

The research prompt that forces AI to cite its sources

True story. Earlier this week I asked AI to compare 2 email platforms. Prices, features, & enough trade-offs to help me pick one.

I had a clean report back in 39 seconds. So beautifully formatted I almost forwarded it to a friend.

Then a source it cited caught my eye.

Dead link.

Tried another & that article was 2.5 years old, its prices long since sailed off into the sunset.

Claude never opened a single site. It faked the report from the bits it already knew.

Same way I faked book reports on the Odyssey in 9th grade.

The fantasy sat shoulder to shoulder with the facts & I couldn’t tell which was which.

So I rebuilt the prompt from the ground up. It has to show its work now (when it can’t, it tells on itself & marks the line unverified).

You are my research assistant. Research the topic below and hand me a brief I can trust. Follow every rule; do not skip the retrieval rules.

TOPIC: [drop your topic here]
PURPOSE: [decision / presentation / just understanding]
DEPTH: [quick scan / standard / deep]

BEFORE YOU RESEARCH
Turn my topic into the 3 specific questions worth answering. Show me the 3 questions first, then answer them.

SOURCING RULES (these are the point of this prompt)
1. Cite inline. Put the real URL in parentheses right after each claim, not in a list at the end.
2. Retrieved only. State a fact only if you actually pulled it from a source in this session. If you can't verify it, write "Unverified:" in front of it. Never invent a URL, statistic, date, or quote.
3. Match the source to the claim:
   - Factual data (prices, specs, dates): the official/primary source is fine. Use the company's own page for its own numbers.
   - Opinion or comparison ("best," "better," "worth it"): independent third-party sources ONLY. A company can never be the source that says it is the best.
4. Triangulate anything that matters. For key numbers, check a second source. If a primary source and a third party disagree, show both and trust the primary for its own facts.
5. Date every source. Flag anything that looks stale.

STANCE
Stay neutral. Assume no preferred answer and do not tailor the verdict to me. If the honest answer is "it depends," say what it depends on.

OUTPUT
Query: [restate the exact question researched]
The 3 questions: list them.
TL;DR: 2-3 sentences.
Findings: one short section per question, each factual claim carrying its inline source. Mark unverified items plainly.
Where sources disagree: name the conflict and which you trust and why.
Verdict: the neutral answer, third-party sources only.
Confidence: rate each major finding High / Medium / Low, tied to whether you retrieved and corroborated it.
Still open: what you couldn't verify.

Why it works: It makes the model cite inline & only from sources it actually pulled up, so you can see which claims are real. It splits the sourcing rule the right way, ask a company for its own price, never for whether it’s the best. When it can’t verify something, it says so instead of making it up.

The Teardown

The viral “grill me” prompt that pressure-tests any plan

There’s a Claude skill on GitHub with over 174,000 stars & its dev, Matt Pocock, is still updating it.

If you’re a dev, you’ve probably seen /grill-me. It’s part of a whole suite of skills he put out.

If you had one guess at what this skill does, you’d probably get it right.

The whole thing is wrapped around a single prompt & that prompt does all the heavy lifting. Which means you can run it on any model.

Saves everybody a bucketful of time & headache before they burn a weekend building the wrong thing, only to get their hearts broken later on.

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer. If a question can be answered by exploring the codebase, explore the codebase instead.

Take that last line out and the prompt is wide open. Point it at a course, a product you can’t position, a decision you keep circling.

Then make it hand you a TLDR, because the grill-me skill talks more than a toddler on a sugar high. Opus especially.

Interview me relentlessly about [my plan or decision] until we reach a shared understanding. Walk down each branch of the decision tree, resolving dependencies one at a time. For each question, give me your recommended answer. Ask one question at a time and wait for my reply. Keep each turn short, and give me a TLDR at the top of any long answer.

This is the meal-planner move from a previous issue. Interview first, answer second.

Via Matt Pocock (mattpocock/skills, GitHub).