Using AI and directing AI are different skills

Hello from the HRBP Lab. If you just joined, welcome, I'm glad you're here. If you are returning - welcome back! This week has been an exercise in balance, focus and prioritizing. Prepping to launch a large, complex and highly visible talent program, managing client leaders' need for advice and counsel and the myriad of fires that pop up along the way - classic HRBP work. I unwind by listening to podcasts. One of my favorites, "The Talk Show with John Gruber" lets me nerd out on all things Apple technology. In the latest episode, John and his guest Adam were talking about how they use AI to create software and they both act as the "Director" - they make choices and apply judgment based on their taste. It got me to thinking about how I use AI in my world as an HRBP and how I can develop the skill of being the director or tastemaker for what AI returns when I prompt.

Much of AI in HR is stuck on using: type the prompt, take what comes back, maybe review a little and move on. Using is real, and it's where everyone starts. But the skill that separates the HRBPs getting hours back from the ones collecting drafts is the upstream one. They decide what the output should be before it exists, and they judge it when it arrives. They direct.

The director doesn't operate the camera

A film director doesn't run the camera or hold the boom, and often can't do either job as well as the crew. The director decides what the shot is for and whether the take works. That's the role: a standard, an audience, and a verdict.

Directing an AI output is the same move. What should this thing be? Who is it for? Is this take actually good, or just competent? The HRBP who directs decides what a strong comp narrative looks like before Claude drafts it, then rejects the take that misses, even when every sentence in it is defensible. The HRBP who only uses accepts whatever arrives, because there's no standard to judge it against.

"Competent but wrong" is the case that matters. A bad draft rejects itself. The dangerous draft is the one that's fluent, plausible, and not quite what the moment needs. Only a director catches that one.

Taste sounds like a luxury. It's the job.

The judging skill has a name: taste. In HR that word can sound like an indulgence, something for design teams. It isn't. Taste is the judgment that knows this manager talking-point will land and that one will put someone on the defensive. You built it the slow way, watching real conversations succeed and fail for years. AI doesn't have it. You do.

Back to that podcast from up top — here's the part that stuck. Adam Lisagor is a creative director; his company, Sandwich, has made commercials for half the tech industry. He's not a programmer. This spring he shipped Hovercraft, a Mac app he built with AI coding tools. What made the app good wasn't the code the AI wrote. It was Lisagor deciding, take after take, what the app should feel like and which versions weren't it. The AI wrote the draft; the director decided whether it was any good.

That second job is the one HRBPs are built for.

You earned the taste by doing the work

In Issue 004 I wrote that I learned to direct the automation by doing the work first. I want to say that part plainly: you can't direct what you've never done. The HRBPs who judge AI output well are the ones who wrote the comp narratives by hand, sat through the calibrations, and watched a talking-point fail in a real room. The work taught the taste. The taste makes you a director. The director is the role the tool can't fill.

It's also why the "AI will replace the juniors" panic gets the order backwards: take away the work and you take away where taste comes from. That's a pipeline question I took up in Issue 004. For this issue the point is simpler. If you've done the work, you already hold the qualification. Most HRBPs just haven't started using it on purpose.

Direct one AI output this week

Not a course, not a new tool. One real task you were going to hand to AI anyway, directed instead of used:

1. Before you prompt, decide what good looks like. One sentence. "A comp message that holds firm without sounding defensive." If you can't name it, the AI can't hit it.

2. Name who it's for. A nervous first-time manager and a skeptical VP need different versions of the same message. Direction starts with the audience, not the topic.

3. Judge the take against your standard, and reject the misses. The first draft is a take, not a deliverable. Ask the director's question: is this actually good, or just competent? Send it back until it clears the bar you named in step 1.

4. Keep the standard, not the draft. The reusable asset isn't the output. It's the "what good looks like" sentence, because that's what you direct with next time. Drafts expire; standards compound.

By Friday you'll have directed one real output: a named standard, a named audience, and at least one rejected take. That last one matters most. The first time you turn down a competent draft, the role shifts.

Before you accept the next draft

This week, when AI hands you something that reads fine, pause before you take it. Ask whether it's good or just competent. If the honest answer is competent, send it back with a note about the job it didn't do.

Competent is what the tool gives you. Good is a decision. Somebody in the room has to make it, and it should be you.

What's the last AI draft you sent back, and what standard did it miss? Hit reply — I read everything that comes in.

-Josh

The prompt I'm using this week

This pairs with the framework above. Run it when you have a draft in hand and you're about to hit accept.

"Act as my director's check. I'm an HR business partner reviewing a draft before it goes out. My standard for this output, in one sentence: [WHAT GOOD LOOKS LIKE]. The audience: [WHO IT'S FOR — role and state of mind]. Here's the draft: [PASTE DRAFT]. Don't rewrite it yet. Judge it against the standard and the audience only. Tell me: (1) where it meets the standard, (2) where it's competent but misses, and (3) the one change that would close the biggest gap. Point to specific lines, not general impressions. If it genuinely clears the bar, say so and stop."

Why it works: Most AI review loops ask the tool to make a draft better, which produces endless polishing with no verdict. This one makes Claude judge instead of generate, against a standard you named before the draft existed. The "competent but misses" category is the whole game — it's the gap a director sees and a user doesn't. And the last line gives the tool permission to stop, which keeps you from sanding a good draft into a worse one.

I've started running this on manager talking-points before they leave my desk. The standard sentence takes longer to write than you'd think. That's the work.

If you try it, tag it #buildtogether and show me what you built — I'll feature the best ones in a future issue.

If you want to go to the source:

• The Talk Show, Ep. 447 — with Adam Lisagor — John Gruber, May 15, 2026 (the Hovercraft build conversation)

• The Extra Time Problem — Issue 001, 2026-04-25

• Feeling like it's Groundhog Day? AI can help. — Issue 002, 2026-05-09

• What only a human can do — even when AI could — Issue 003, 2026-05-16

• Where do tomorrow's HRBPs come from? — Issue 004, 2026-05-23

• Nobody owns the seam — Issue 005, 2026-05-30 (VERIFY slug before send)

If someone you know is navigating the same question, forward this their way. New here? Subscribe at thehrbplab.ai to get the next one.