What's closest to your day-to-day?

The gap between a vague ask and a usable draft

Most disappointing AI writing comes from under-specified requests. The model isn't bad at writing, it just didn't know what you actually wanted.

The second version takes 15 extra seconds to type and saves you a full rewrite. Specificity is the entire skill.

Three things worth specifying every time

• Audience, executives, your team, a client, all read differently
• Length: "keep it to 3 sentences" changes the output more than almost anything else
• What to leave out, telling it what NOT to include (excuses, jargon, a specific topic) is as useful as saying what to include

Give it the "so what," not just the data

The failure mode in analysis work is pasting numbers and getting back a bland restatement of numbers you already had. The fix: tell it what decision the analysis is for.

The pattern worth learning

Ask for the interpretation, not the description. "What changed" is something you can already see in a spreadsheet. "Why it probably changed, and what I should do about it" is the part worth asking an AI for.

Turn ambiguity into a structure it can work with

PM work is full of half-formed lists and shifting priorities. AI is genuinely good at imposing structure on mess, but only if you hand over the actual mess, not a cleaned-up summary of it.

Why pasting the mess works better

A cleaned-up summary has already lost the details that matter, dependencies, half-finished thoughts, things you weren't sure fit. The raw version gives the model more to work with, the same way a messy first draft gives a human editor more to work with than a one-line summary.

The tone instruction matters more than the content

Client-facing writing lives or dies on tone, not facts. AI defaults to a slightly generic, overly formal register unless told otherwise, which is exactly the tone that makes an email feel like a form letter.

The habit: describe the relationship, not just the message

"We have a good relationship" or "this client is difficult and needs formal language" changes the output more than any content instruction. Relationship context is the missing ingredient in most bland client drafts.

Practice round 1: write your real request

Below, write an actual request from your real work this week, not a test question. Use what you just learned: specify audience, length, and what to leave out.

The technique that matters more than any single prompt

Your first response back was probably 80% right, not 100%. That's normal and expected, the skill isn't writing the perfect first prompt, it's knowing how to refine.

Refining well looks like this

• "That's close, but make it shorter and cut the last paragraph"
• "Good, but this doesn't sound like something I'd actually say, make it more direct"
• "Keep the structure, but the second point needs a specific example"

Each of these takes 5 seconds to type and gets you further than starting over with a longer, more "perfect" prompt would have.

Practice round 2: refine your own draft

Now ask for one specific change to what you got back in round 1, shorter, different tone, a missing detail, anything real.

You just tried the real thing, here's the rest

Those were genuine AI responses, not a demo. The rest of this lesson, all 5 lessons in this path, and everything else in AI Learning Loop unlocks with a paid membership.

The habit that protects you

Always ask for a first draft, never treat the output as final. The biggest time-loss isn't using AI too little, it's sending output without reading it closely enough to catch a wrong name, a made-up statistic, or a tone that doesn't match the relationship.

A real cost of skipping this: a wrong client name in a "personalized" email reads worse than a generic one, it signals nobody actually looked at it.

Putting it together: a request checklist

Before you send any work request to AI, you now have four things worth checking:

1. Did I specify audience and length?
2. Did I paste the real, messy input rather than a cleaned-up summary?
3. Am I ready to refine, not expecting a perfect first answer?
4. Will I actually read the output before it goes anywhere?

This is the entire skill. Everything else is practice.

What matters most in a request that isn't working?