The Second Ask

Watch somebody use AI badly for ten minutes and you'll see the same thing happen every time.

They write a prompt. They get back something mediocre. They stare at it. And then they delete the whole thing and write a slightly different prompt, from scratch, in a new window.

Then they do it again. And again. And around attempt five they conclude the tool is overrated and go back to doing it by hand, having spent forty minutes proving something that isn't true.

The problem isn't the prompt. It's that they treated a conversation like a vending machine. Put in coins, get wrong snack, kick machine, try different coins.

The best operators I know write worse first prompts than the people who read prompt guides. Genuinely worse. Sloppier, shorter, less structured. They just never accept the first answer, and they know exactly what to say next.

That second move is where all the value is, and it's the part nobody writes about, because "how to recover a bad output" makes for a much less exciting listicle than "the ultimate mega prompt."

So here's the recovery kit. Six moves. Each one fixes a specific failure, and knowing which one to reach for is most of the skill.

Before the moves, the diagnosis, because reaching for the wrong tool is why people give up.

When output comes back bad, it's bad in one of these ways. It's generic. It's confidently wrong. It's the right content in the wrong shape. It's technically correct and lifeless. It's answering a slightly different question than you asked. Or it's fine but you can tell it's playing safe.

Different diseases. Different medicine. Match them and your hit rate goes up enormously.

Move one: make it interrogate you

This is the one that fixes about half of everything, and if you take one thing from today take this.

When output is generic, the cause is almost always that you didn't supply enough specifics and the model filled the gap with the average of everything it's ever seen. That's not a flaw, that's what it has to do when you leave a hole.

So instead of writing a longer prompt, hand the question back.

Then actually answer them.

What happens next is slightly uncomfortable and very useful. The questions are usually good, and about a third of the time you can't answer one of them, which means you just discovered a hole in your own thinking that you were about to paper over with confident copy.

The one at a time part matters more than it looks. Five questions in a block get answered lazily, in a batch, with the last two getting a shrug. One at a time and each answer sharpens the next question.

This move turns a two minute mediocre output into a six minute good one. That's the actual trade and it's a bargain every time.

Move two: ask for the version it didn't give you

For output that's competent and dead. Nothing wrong with it. Nothing right either. Reads like a brochure for a company that doesn't exist.

That's the model playing the odds, which is its job. So change the odds.

The second sentence is what makes this safe to use. You get the braver draft and a map of where it's shaky, so you can keep the energy and check the risky bits rather than throwing out both.

Variant for when the writing is fine but boring: ask for three openings that would make a skeptical reader keep going, then say which one you'd bet on and why. The reasoning is often more useful than the openings.

Move three: make it argue with itself

For output that's confidently wrong, or plausible in a way you don't trust.

Models are extremely good at generating a persuasive case and not naturally inclined to stress test it unless asked. Ask.

This is the move I use before any decision with money attached. It's caught me twice on things I was about to spend real money on, both times because the counter case surfaced a cost structure I'd waved past.

Related move for factual claims: pick the three specific claims that would most damage the argument if false, and tell me how to check each one. You get a verification list instead of a vague warning, and you can actually work through it.

Move four: separate the content from the container

For output where the substance is right and the shape is wrong. The points are good, the format is a wall of bullets, or an essay when you needed a table, or five paragraphs when you needed three sentences.

Do not re prompt. The content is already right and you'll lose it.

Or: as a one page brief with a two sentence summary at the top. Or: as an email to a busy person who'll read the first line and decide. Or: cut this to a third of its length without losing a single distinct idea, and tell me what you cut.

That last one is the most useful editing prompt in existence and I use it constantly. "Tell me what you cut" is the important half, because sometimes it cut something you wanted and now you know.

Move five: give it something real to work from

For output that's plausible but not yours. Sounds like a competitor. Sounds like anyone.

The fix isn't a better description of your voice. It's an example.

The two step matters. Asking directly for imitation gets you a caricature, usually your three most obvious tics turned up to eleven. Making it articulate the pattern first, then apply it, gets something that holds up past the first paragraph.

Keep three or four pieces in a file for this. Actual work that landed. It's the highest leverage input you own and almost nobody maintains it. If you already run something like Rize and can see where your hours actually went last quarter, the work that took longest is usually the work worth keeping as an example.

Move six: ask what you should have asked

For when nothing's working and you can't tell why.

This one feels like giving up. It isn't. About a quarter of the time the reframe is genuinely better than what you were going for, because you were solving the problem you'd named rather than the one you had.

I've had this move tell me I was asking for a landing page when the actual problem was that the offer didn't make sense. That's not a prompt improvement. That's a business insight, delivered for free, because I admitted I was stuck.

Abstract moves are easy to nod at and hard to use, so here's an actual sequence from last month, compressed.

I needed a follow up email for people who'd downloaded something and gone quiet. First prompt was lazy on purpose: "write a follow up email to leads who downloaded my guide and didn't reply."

What came back was the email you've received four hundred times. "I wanted to circle back and see if you had any questions." Generic, because I'd given it nothing.

Move one. Asked it to interrogate me first, one question at a time. Third question was "what specifically were they trying to solve when they downloaded it, and how do you know?" I didn't have a good answer. That's the moment the exercise paid for itself, because the reason my follow ups were weak was that I'd never actually established why anyone downloaded the thing.

Answered the five questions properly. Second draft was substantially better and still felt like a template.

Move five. Pasted in two emails I'd written that got replies, asked it to identify the pattern before rewriting. It came back with something I hadn't consciously known about my own writing: my emails that work open with a specific observation about the recipient's situation and never mention the thing I sent them until the third line. That's true. I'd never articulated it.

Third draft, using that pattern, was close.

Move four. Content was right, length was wrong. Cut to a third, tell me what you cut. It cut a paragraph I wanted, so I put that one back and kept the rest of the trim.

Move three, briefly. Asked what would make this email annoying to receive. Got three answers, two of which I disagreed with and one of which was correct: the ask at the end was vague enough that replying required effort. Fixed that.

Four moves, one thread, maybe twelve minutes. The result outperformed the previous version by a wide enough margin that I stopped testing.

Nowhere in that process did I write a better first prompt. The first prompt stayed lazy. The work happened after.

Worth naming, because these get taught constantly and they waste your afternoon.

Telling it to try harder. "Be more creative." "Make it better." "This isn't good enough." These do almost nothing, because they contain no information. You've expressed dissatisfaction without specifying the axis. Every one of the six moves works because it names a direction.

If you catch yourself typing a sentence that would be useless feedback to a junior employee, it's useless here too. "Make it pop" has never improved anything in the history of work.

Stacking adjectives in the first prompt. The prompt guide industry runs on this: write "you are a world class expert copywriter with twenty years of experience" and watch quality improve. It mostly doesn't, and it definitely doesn't compared to spending the same effort on the second message.

Persona framing has some effect at the margins. It is nowhere near as useful as an actual example of the work you want, which is why move five beats any amount of role assignment.

Rewriting the same prompt five different ways. This is the vending machine behaviour from the top of the piece. If attempts one through three came back wrong in similar ways, attempt four phrased differently will come back wrong in a similar way, because the missing thing is context you haven't supplied rather than words you haven't found.

The tell is when you notice you're editing your prompt rather than responding to the output. Stop, and use move six.

Here's what actually changes if you internalize this.

You stop treating the first output as the deliverable and start treating it as a probe. You said something vague, you got back the average, and now you know where the vagueness was. That's information. Use it.

Practically: stop opening new chats. The single most common bad habit I see is people abandoning a thread the moment it goes sideways. Everything you've built up in that conversation, the context, the corrections, the examples, all of it dies when you start over. The recovery moves only work because there's something to recover.

Second: keep the moves visible. Not memorized, visible. A text file, a snippet manager, a note pinned in whatever you use. You will not remember them under pressure at 4pm and that's exactly when you need them.

Third, and this is the one that took me longest: notice when you're on attempt four. Four attempts means the tool isn't the problem. Either you don't actually know what you want, which move six will surface, or this is a task the model is bad at and you should do it yourself.

Knowing when to stop is a skill too. Most of the frustration people have with these tools comes from pushing on the fifth attempt at something that was never going to work, when the first attempt at something adjacent would have worked fine.

Six moves. Print them, pin them, use the first one today.

Jordan

The AI Newsroom is written by Jordan Hale. This issue contains affiliate links to tools I actually use. If you sign up through them I may earn a commission at no extra cost to you.