Useful AI output starts with a useful ask: audience, context, tone, length, constraints, and what good looks like.
Previously On ... She opened all three tools. Typed the same thing into each. Rated them on the butterfly clip scale. Picked her favourite.
She opened all three tools. Typed the same thing into each. Rated them on the butterfly clip scale. Picked her favourite. Survived entirely. (Episode 1: On Wednesdays We Use AI)
On This Episode... The one in which she types something vague into ChatGPT, gets back the world's most useless paragraph, channels her inner David Rose.
The one in which she types something vague into ChatGPT, gets back the world's most useless paragraph, channels her inner David Rose demanding very specific things, tries again with actual context, and watches AI produce something she'd genuinely send to her VP. She discovers that prompting isn't coding. It's delegation. And the Spice Girls were onto something.
I couldn't help but wonder...
...why AI reads my mind sometimes and completely ignores me other times.
Last week I asked it to help me draft talking points for a meeting. I got back something so generic it could have been a LinkedIn motivational poster printed on a throw pillow. The next day I asked it to help me write a project update for my director, told it who my director is, what she cares about, what the project actually does, and what I needed her to approve. It came back with something I barely had to edit.
Same tool. Same Tuesday afternoon. Completely different results.
I spent an embarrassing amount of time thinking the tool was inconsistent before I realized the variable was me. What I was giving it determined what I got back. Almost without exception.
Tell Me What You Want, What You Really Really Want
Why AI gives you garbage and what to do about it.
You know the party game where one person describes something and the other person has to draw it based only on the description? You say "draw a house" and you get a child's crayon box with a triangle roof. You say "draw the Victorian from Practical Magic with the wraparound porch, the widow's walk, and the overgrown garden" and suddenly the picture looks like something.
That's prompting. That's the whole concept. You're describing what you want to someone who will execute it precisely and literally, but who cannot read your mind, doesn't know your context, hasn't been in your meetings, and won't always ask a clarifying question before charging ahead.
The Spice Girls had it figured out decades ago: tell me what you want, what you really really want. (Also, decades?! Yikes, that is a jagged little pill to swallow.) They weren't asking for vague gestures toward a general desire. They wanted specifics. AI wants the same thing. It will zig-a-zig-ah all day long if you let it, but it produces useful work only when you tell it exactly what you need.
The David Rose School of Prompting
How to actually fold in the cheese.
If you've watched Schitt's Creek, you know David Rose is impossibly, almost pathologically specific about what he wants. The wine has to be a very specific wine. The sweater has to drape a very specific way. The brand aesthetic must be "a general store but make it specific." He's exhausting in life but he would be absolutely brilliant at prompting AI.
Because specificity is the thing. Not fancy language. Not special formatting. Not knowing technical terms or secret codes. Just being David Rose levels of particular about what you're asking for.
And remember when Moira tells David to "fold in the cheese" and he's standing in the kitchen screaming "WHAT DOES THAT MEAN?!" That's your AI when you type "write me an email about the project." It's trying to do what you asked. But you've given it "fold in the cheese" when what it needs is "take the spatula, scrape along the bottom of the bowl, bring it up and over gently, rotate the bowl, repeat." (Those who know me know I 100% relied on AI for that cooking description, because I do not cook.) AI can't read your mind. What's in your head stays in your head until you type it out.
I learned this the hard way. I used to type things into AI the way I'd type a Google search: short, vague, hoping it would figure out what I meant. "Help me write a project update." "Draft talking points." "Summarize this." And I'd get back something that was technically correct and completely useless, like ordering "a coffee" and getting handed a lukewarm medium roast with no milk when what you actually wanted was an oat milk latte with an extra shot.
The shift happened when I started talking to it the way I'd delegate to a smart new hire on their first week. Not dumbing things down, but giving context. Who is this for? What do they care about? What's the tone? What does success look like? How long should this be? What should it absolutely NOT include? What am I actually looking for here? Have we done anything similar before that it can use as a reference?
It's Britney, Bitch
Same task. Two approaches. Watch what happens.
Picture this. Silicon Valley, 2026. You have a 12-page policy change document that needs to get to your stakeholders before a meeting. You paste the full document into the chat and type:
The vague version (what you'd be doing without ):
"Summarize this policy change for my stakeholders."
What AI gives you back:
"The policy change introduces new requirements for compliance reporting. Key stakeholders should be aware of updated timelines and documentation requirements. The changes aim to streamline processes and improve accountability across the organization. Teams should review the full document for specific details relevant to their department."
Four sentences of absolutely nothing. No specifics. No action items. No indication it understood who your stakeholders are or what they'd do with this information. It's the equivalent of Ross screaming "PIVOT." Technically a word, communicating nothing useful, increasingly frustrating for everyone involved.
The specific version (what you type after reading this Episode):
"I need to summarize a policy change for 6 senior managers who have 2 minutes to read this before a meeting. They care about: what's actually changing, when it takes effect, what their teams need to do differently, and whether it affects budget. They don't care about the history of why this changed or the regulatory background. Tone: direct, no fluff, bullet points are fine. 150 words max."
What AI gives you back:
What's Changing: Quarterly compliance reports now require sign-off from both the team lead AND finance partner (previously just team lead).
When: Effective Q3 2026. First dual-signed report due September 30.
What Your Team Needs to Do:
• Identify your finance partner by July 15
• Add them to your reporting workflow
• Build in 3 extra business days for the additional review cycle
Budget Impact: No direct cost. But factor in ~2 hours of finance partner time per quarter per team.
Action required by you: Confirm your finance partner assignment by July 15.
That second output took about 90 seconds to write the prompt and maybe 45 seconds to review. You change one detail, hit send to your stakeholders, and move on with your life.
It's Britney, bitch. That's the feeling when a prompt comes back right. That completely unbothered confidence of knowing exactly what you asked for and getting exactly what you needed.
40%
Harvard and BCG studied 758 consultants using ChatGPT-4. For tasks inside AI's "jagged frontier," consultants using AI completed work over 25% faster and produced output rated over 40% higher quality. For tasks outside that frontier, some users did worse because they trusted polished wrong answers. Same tool. The difference was judgment, context, and knowing when to verify.
BCG x Harvard Business School, 2023 · 758 consultants
And here's what's interesting about that. The skills that made the difference weren't technical. They were the same ones you use to delegate, to brief a new team member, to tell someone exactly what you need without micromanaging. Ethan Mollick, the Wharton professor everyone in AI cites, called it:
The skills so often dismissed as 'soft' turned out to be the hard ones.
Ethan Mollick, One Useful Thing
Turns out the woman who writes a thorough project brief is better at prompting than the engineer who types three words and expects the machine to read between the lines.
So You Don't Pull a Cher
Quick definitions so you can use these words confidently in a meeting without accidentally arguing that it doesn't say RSVP on the Statue of Liberty.
Prompt What you type into the chat box. Remember Elle Woods' Harvard application video essay? She didn't just say "let me in." Read full definitionShow less
What you type into the chat box. Remember Elle Woods' Harvard application video essay? She didn't just say "let me in." She was specific, personal, tailored to her audience, and impossible to ignore. That's a good prompt. It's not code. It's not a spell. It's you being clear about who this is for, what you need, and what "good" looks like. You've been writing prompts since last week. You just didn't know it had a name.
Context The background information that makes AI understand YOUR situation, not just the generic version. Remember Miranda Priestly's cerulean speech? Read full definitionShow less
The background information that makes AI understand YOUR situation, not just the generic version. Remember Miranda Priestly's cerulean speech? Andy thinks she's just wearing a blue sweater. Miranda traces the entire history of that specific shade through Oscar de la Renta, Yves Saint Laurent, eight different designers, and a clearance bin to explain why "blue" is never just "blue." That's what context does for AI. When you say "this is for senior managers who have 2 minutes and don't care about the backstory," you've just given it the cerulean speech. Without context, AI gives you generic blue. With it, you get cerulean.
Token A chunk of text the model processes. You do not need to obsess over tokens yet; just know that more context has a cost and tools have limits. Read full definitionShow less
A chunk of text the model processes. You do not need to obsess over tokens yet; you just need to know that more context has a cost and tools have limits. If a chat starts acting forgetful, vague, or suddenly less aware of what you already said, it may be because the conversation has more material than the tool can comfortably hold at once.
