Most "best prompts" lists are collected, not tested. This one only keeps what held up when it was actually run.
Most “best AI prompts” content is collected, not tested: someone finds a long, impressive-sounding prompt, screenshots it, and it spreads because it looks like effort, not because anyone checked whether it beats a plain instruction. This guide pulls nine prompts from real, identifiable public sources and runs every one of them on a stand-in task before deciding whether it belongs. Six make the cut, organized by the job they solve rather than by tool or role: getting AI to ask what it doesn’t know, arguing back at your own decision before you commit, getting past your first three ideas, turning a messy meeting or inbox into something you can act on, and rehearsing a conversation you can’t wing. The other three are named and explained, including the elaborate “expert persona” prompt style sold in most prompt packs, which tested no better than a plain, well-constrained instruction.
Picture a marketing manager twenty minutes before a client call, no deck started yet, pasting a four-hundred-word “God-tier ChatGPT prompt” she screenshotted from a stranger’s post on X the night before. It promises a killer intro slide in one shot. What comes back is confident, generic, and wrong about the client’s industry. She doesn’t notice until she’s already presenting, but the shape of it will feel familiar.
That’s the problem with most “best prompts” content, including the versions with a hundred thousand upvotes behind them. They’re collected, not tested. Someone finds a prompt that sounds impressive, it spreads because it’s long and specific-sounding, and nobody checks whether it actually beats a plain instruction.
For this guide, nine prompts genuinely popular in identifiable public places, Reddit threads, a productivity creator’s own site, a researcher’s published writeup, got run on a stand-in task before earning a spot here. Six held up. Three didn’t, and section six says exactly why, because a shorter list you can trust beats a longer one you have to guess about.
Start with the prompt that solves the problem in that opening scene: getting the AI to say what it’s missing before it writes something confidently wrong.
The version that shows up most often, including in a widely read Tom’s Guide how-to, is one line added to the end of whatever you’re already asking for: “Ask me any contextual questions about my request before you give me an answer.”[1] It works. Run on a stand-in brief, “write a one-paragraph client intro for a company I haven’t described yet,” it came back with the three questions that actually mattered: industry, tone, and what the client cares about, instead of guessing all three and getting one wrong.
Adapted from the version reported in Tom’s Guide’s own testing. [1]
Here’s the part most versions of this tip leave out, and it’s the difference between the prompt working and just feeling like it worked. Answer the AI’s questions in the same chat thread and it’s already halfway through a plan built around your original, vague prompt, so the answers get bolted on rather than built in. Tested both ways on the same brief: answering in-chat produced a serviceable draft. Folding the answers straight into a rewritten version of the original prompt and resubmitting it produced a noticeably tighter result, with no hedging language trying to cover two tones at once.
When a prompt asks you clarifying questions, put the answers back into the original prompt and resubmit it. Don’t just reply underneath.
This one is worth keeping close if you’re the person whose “quick AI draft” becomes someone else’s starting point: an HR coordinator drafting a policy summary for legal, a marketer briefing a designer. The extra fifteen seconds of rewriting the prompt is cheaper than the round trip when the first version misses the brief. For a fuller way to pressure-test any prompt before trusting what it hands back, see our guide on how to test your AI prompts before you trust the output.
Getting AI to ask better questions fixes one common failure. It doesn’t fix the other one: an AI that agrees with whatever you already believe, just in more polished sentences. Ask “is this a good idea?” and you’ll usually get “yes, and here are three reasons,” because agreement is the easier response for these models to produce.
The fix that shows up again and again, independently, across prompt-engineering write-ups, productivity newsletters, and Reddit threads, is one of the same handful of prompt patterns worth learning once and reusing everywhere (we map the others in our guide to the prompt patterns behind every good AI answer). There’s no single person who invented this particular one. It’s a two-step instruction: build the strongest version of your position, then attack that version as hard as possible.
A widely repeated prompt pattern with no single identifiable origin, tested here on a stand-in decision.
Tested on a stand-in decision, “we should cut our email newsletter from weekly to monthly to reduce fatigue,” the steelman came back solid and mostly expected. The counterargument was the useful part. It named a real risk: a monthly cadence can break the habit loop that gets people to open on a predictable day, and the “fatigue” the decision was meant to fix might be a subject-line problem rather than a frequency one. That’s a genuine consideration worth weighing, not a manufactured objection for the sake of having one.
One honest limit: this works best on a decision you’re fairly confident about. Run it on something you’re already unsure of and the “attack” step mostly produces noise dressed up as insight, since there’s no strong position yet to steelman in the first place. A couple of quick checks before you run it:
It’s worth reaching for specifically when a decision is about to leave your desk and land on someone else’s: a manager recommending cutting a program, a finance lead framing a budget ask before it reaches leadership. Arguing against yourself first is cheaper than hearing the argument for the first time in the room.
Idea generation is the other place a default prompt quietly under-delivers. Ask “give me five ideas for X” and you’ll get five safe, generic ones, roughly the same five most people would get, because the model reaches for the most statistically likely answer first.
Ethan Mollick, the Wharton professor who has run and written up a series of real experiments comparing AI and human creative output, found something specific: quantity beats a single polished attempt, because the ideas are meant to be judged afterward, not while they’re being produced.[2] His own writeup describes prompting for roughly ten ideas at a time, repeating that a few times, then evaluating the full batch, rather than asking once and taking whatever comes back.
Structure adapted from Ethan Mollick’s published testing of AI idea generation. [2]
Run on a stand-in brief, ideas to re-engage a dormant email list, the first ten were a mixed bag: three genuinely reusable, a few generic (“send a discount code”), one that only makes sense if you enjoy customer complaints. A second batch, run right after with no new instructions, produced almost no overlap with the first, which is the actual point: quantity buys variety that a single “give me five good ones” ask doesn’t. The evaluation step at the end, ranking each for feasibility and originality, is what turns fifteen mixed ideas into three worth actually pitching.
This is the one most worth handing to whoever owns a recurring “we need fresh ideas” task: a content marketer rebuilding a campaign brief, an L&D lead planning next quarter’s training topics, an operations person looking for a workaround before escalating. It’s less useful for a decision that already has two clear options; the previous prompt is built for that instead.
Say an HR coordinator is sitting with forty minutes of scrawled notes from a benefits-vendor call on a Wednesday afternoon, three action items buried somewhere in there, and a 4pm deadline to send a recap before everyone forgets what was agreed. That’s the ordinary, unglamorous case most prompt roundups skip past in favor of something flashier.
Jeff Su, a productivity creator whose own site walks through this exact use case, builds his version around giving the model a fixed output shape instead of just asking it to “summarize.”[3] The shape matters more than the wording: a one-line summary, the decisions that were actually made, and a table of action items with an owner and a date, and it should say so explicitly when an owner isn’t clear instead of inventing one.
Adapted from Jeff Su’s structured-summary approach on jeffsu.org. [3]
Tested on a stand-in transcript with one action item that had no assigned owner in the raw notes, the prompt did what it was told: it flagged the item as unassigned instead of guessing a name, which is the behavior you actually want from something you’re about to forward to a room of people who’ll trust it at face value.
This is worth having ready for anyone whose week is mostly back-to-back calls: HR running intake meetings, a finance lead sitting through budget reviews, a marketer coming out of five client calls in a day. The version worth keeping is the one that admits what it doesn’t know, not the one that fills every cell in the table just because the table has cells.
Some conversations only happen once, or once every few years, so you can’t learn to handle them well through repetition the way you can with email. Interviews, negotiations, and the performance review you’ve been dreading all fall into that category, which is exactly where a good rehearsal prompt earns its place.
For interview prep, the useful version isn’t “give me ten interview questions,” which produces a script you read past rather than a rehearsal. It’s closer to the approach behind Jeff Su’s job-search prompts: put the model in the interviewer’s seat, ask one question at a time with real follow-ups, and save the harsh feedback for the end.[3]
Adapted from Jeff Su’s interview-prep approach, tested here on a stand-in internal-transfer interview.
Say a finance analyst is two days out from a panel interview for an internal transfer. Run through this and the model follows up the way a real interviewer does, on the vague part of an answer, not the part rehearsed well. The critique at the end was specific in testing: it named a moment where a project’s outcome was described without saying what the person had personally decided, exactly the gap a real panel would notice too.
For negotiation, there’s no single identifiable source for the version that circulates, but the shape is consistent enough across public write-ups to use anyway: a likely fallback position, two or three realistic opening positions, and the questions the other side will probably ask that you don’t have a strong answer to yet.
A common prep pattern with no single identifiable origin, tested here on a stand-in vendor renewal.
Tested against a stand-in vendor renewal, it correctly flagged an unprepared spot: no ready answer for why the account was even considering switching, which is close to the first question any vendor actually asks. That’s the value, not the opening positions themselves, which you could draft yourself, but the gap it surfaces before someone else does.
Whoever owns the actual conversation should be the one running this, not their manager on their behalf: the internal-transfer candidate, not the current boss; the person on the vendor call, not whoever approved the budget. The prep only helps if the person under pressure in the room is the one who rehearsed it.
The prompt packs sold on marketplaces and pinned to “ultimate ChatGPT prompt” threads mostly follow the same recipe: a long persona (“You are a world-class strategist with 20+ years at Fortune 500 companies, an award-winning track record, and expert-level judgment…”) stacked in front of the actual instruction. A representative version of that style, not copied from one specific seller, was run here against the same dormant-email-list brainstorm from section three, with identical constraints on both sides: ten ideas, prioritize variety, evaluate for feasibility afterward.
The outputs were close enough that telling them apart blind wouldn’t have been easy. The persona version opened with a confident paragraph about “leveraging deep industry expertise,” then landed on close to the same ten ideas. That lines up with the underlying research: when researchers tested basic prompts against ones giving the AI an expert persona, including named famous experts, no single approach clearly won.[2] The elaborate framing didn’t hurt. It didn’t help enough to justify four extra paragraphs of setup either.
A prompt gets better when you add a constraint the model has to satisfy, not when you add adjectives about who it’s supposed to be.
Two more didn’t make the cut, worth naming rather than quietly dropping. “Roast my resume,” a trend with no single identifiable source, produced sharp-sounding but generic notes on a stand-in resume with no job description attached: the same complaints (quantify your impact, cut the objective statement) any resume blog would give. Paired with an actual job posting, the same prompt got specific about which bullet points wouldn’t survive an ATS scan. The anchor was missing, not the prompt.
The third is smaller: answering the clarifying-questions prompt from section one inside the same chat, rather than folding the answers into a rewritten prompt, tested measurably weaker on the same brief. That’s why the rewrite version is the one that made the cut.
A few honest signs a prompt is underdelivering, worth checking before trusting any list, including this one:
These six don’t remove the judgment call. They move it: to a draft instead of a blank page, a counterargument instead of silence, a flagged gap instead of a guess.
| What the prompt does | What you still own | How to check it |
|---|---|---|
| Surfaces the counterargument or missing detail you’d otherwise skip | Deciding whether that changes the actual decision, or just adds noise | Ask whether it named something specific and checkable, not just “consider the risks” |
| Generates volume fast: ideas, questions, a draft structure | Picking the two or three worth acting on, and rejecting the rest without guilt | Could you defend the pick to a colleague in one sentence, not five |
| Flags what it doesn’t know instead of guessing | Actually answering, in writing, before the work goes anywhere | Reread the final version for confident language covering a gap that was never actually filled |
These prompts speed up the first pass. None of them replace the read-through before something goes out.
One more thing worth checking before building a habit around any of these: whether a prompt is worth keeping is different from whether it worked once. Use is the easy signal: did it produce something useful the first time. Persistence is harder to fake: are you still reaching for it a month later, unprompted. The real test is impact: did the output actually change what got sent, decided, or said in the room. The clarifying-questions rewrite and the interview roleplay are the two prompts here that pass all three so far, a better recommendation than anything said about them the first time they were tried. For the formula behind why a well-constrained prompt outperforms a vague one, see our 4-part prompt formula guide.
Pick one of the six above, whichever matches something on your desk this week, not the one that sounds most impressive. Run it once. If it’s still in use a month from now without this article open in another tab, it earned its place. If not, it was a good screenshot and nothing more.
Every one of these is a plain-language instruction, not a ChatGPT-specific feature, so all six work in Claude, Gemini, and Copilot with no changes needed. The one place you’ll notice a difference is the clarifying-questions prompt: some models are more willing than others to actually stop and ask instead of answering straight through anyway. If that happens, add “Do not answer until you’ve asked your questions” to the end.
Because phrasing feels controllable, and a long, ornate prompt feels like more effort went in, which reads as more likely to work. The research behind the brainstorming prompt in this guide found a basic prompt and one dressed up with an expert persona produced results close enough that no approach clearly won. What actually moved the results was a real constraint, a budget, an audience, a required format, not better adjectives.
No more than using a template someone else designed for a slide deck. The prompt is the instruction, not the output. What still has to be yours is the judgment call about which of the ideas, drafts, or counterarguments it produces are actually worth using, and that’s the part these prompts are built to hand back to you, not take away.
Less than you’d think for the structure, more than you’d think for the content you feed in. The clarifying-questions and steelman prompts work close to as written for almost any role. The meeting-summary and interview-prep prompts need your real context: an HR coordinator’s benefits call and a finance analyst’s panel interview need the same shape but different specifics, so the structure travels and the details inside it don’t.
Bookmark it for now. But if more than two of these become a regular habit, it’s worth moving them into a shared doc or an actual prompt library instead of re-finding this article every time. We cover how to set one up, including which fields are worth tracking, in our guide to building an AI prompt library for your team.
Every prompt kept in this guide was run at least once on a stand-in task before being included, and the three cut prompts are named rather than quietly dropped. The clarifying-questions technique is checked against Tom’s Guide’s own reporting, the idea-generation approach against Ethan Mollick’s published research writeup on AI and creativity, and the meeting-summary structure against Jeff Su’s own productivity site, all fetched and confirmed directly. The steelman and negotiation-prep prompts are patterns that show up independently across many public sources with no single identifiable inventor, and this guide says so rather than inventing a source.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.