How to Take Your AI Program Beyond Prompts
Core Shift: Prompt libraries improve consistency, but workflow-based skills create feedback loops that steadily improve People decisions.
Hiring Loop: AI hiring workflows connect planning, interviews, and later performance data to reveal which signals predict success.
Diagnostic Test: Evaluate every People process by checking inputs, repeatability, downstream artifacts, and whether outcomes return for improvement.
Human Judgment: Compadre expands compensation access while keeping recommendations and final decisions with accountable human professionals.
Leadership Choice: The biggest gains come from redesigning work, integrating data, and developing practitioners who think in workflows.
A lot of AI programs inside a People function start the same way. Someone found a prompt that produced a decent job description, someone else found one that summarized engagement survey themes, and within a few weeks there was a shared document.
Practitioners learned the prompts and felt faster, adoption metrics moved, and after two years of hand-wringing about whether HR would be left behind, that felt like progress. Then six months passed, and somebody asked where the value actually landed.
Brad Wilkins, vice president of People and Organization at Cognite, lived that story inside his own function and went looking for the value. Not very much had accrued, and the reason had nothing to do with the quality of the prompts.
Create a Free Account to Keep Reading—and Keep Leading Smarter
Unlock this piece and join a community of forward-thinking leaders discovering tools, playbooks, and insights for thriving in the age of AI.
Have an account? Log In
What a Prompt Library Can't Accumulate
A prompt is one instruction that produces one output. What a prompt library accumulates is consistency on the input. What it can't accumulate is evidence about the output, because nothing routes the result back.
"Prompts are the opening move," Wilkins said. "They are not the unit of work."
The unit of work is a skill, borrowing the term from the tooling his team builds on. It refers to something that pulls inputs from several systems, runs a consistent process across them, produces a defined artifact, and improves as outcome data comes back to it.
Hiring, Run Both Ways
Run hiring on prompts and you get a good job description in 90 seconds instead of 40 minutes. A manager describes the role, the practitioner runs the saved prompt, and a solid draft comes back. Do it 200 times and the final draft is exactly as good as the first, which is the problem.
The skill version starts earlier and ends much later. At Cognite, intake starts with a structured interview with the hiring manager. That conversation is cross-referenced against finance’s headcount plan and the team’s performance distribution. Inconsistencies surface before anyone writes a word.
Peer conversations are synthesized into one clear hiring need. Then the supporting materials come together, including an anchored rubric for interviewers to score against. During the interviews, the workflow flags things like question quality and talk-time ratio.
Then comes the part that matters. Six months after a hire starts, the system looks back at the interview scores and compares them with how that person actually ramped up. Over time, it can show which managers are good at assessing certain competencies and which interview signals actually predict success.
The 200th hire runs through a more accurate loop than the first one did. That is the whole distinction, and no prompt produces it however carefully it's written.
Wilkins picked hiring deliberately. It's where organizations have some of their messiest data and the biggest differences in manager behavior. That's also why compounding is worth more in a People function than in engineering, where the work is already instrumented.
Four Questions
Wilkins' definition works as a test. For anything your function has built, ask these four questions:
- Does it pull inputs from more than one system?
- Does it run the same process every time rather than depending on who is driving?
- Does it produce an artifact someone downstream depends on?
- Does anything route the outcome back?
Answer no to the fourth and you have a prompt library with more steps. This creates a ceiling.
Wilkins thinks prompt-anchored organizations level off around 10 to 20% productivity improvement, his own estimate rather than a published benchmark. But its shape matches PwC's 2026 AI Business Predictions, which argues that technology delivers about 20% of an initiative's value and the other 80% comes from redesigning the work.
"Had I known this earlier, I would have skipped the months of prompt-library curation that everyone in our field has been doing and gone directly to skill design," Wilkins said. "I would have invested in the practitioners who could think in workflows rather than the ones who were good at writing clever prompts."
He hasn't torn down what he built, and he still argues for treating a prompt library as an organizational asset. Prompts are necessary, but they run out, and the trouble starts when the library becomes the destination rather than the first month of a longer build.
The Payoff Is in the Workflow
Tony Castellanos, head of people at Nextdoor, has overseen the creation of a compensation service that shows what changes when the unit is a workflow. It plugs into the company’s compensation architecture and reads across the employee base. Recruiters and People business partners can access that knowledge directly instead of having to go through the compensation team. They call it Compadre.
A recruiter can ask how a proposed offer compares with the company’s current pay bands. The service also shows how it compares with the pay of people already in the company, without identifying them.
What it will not do is analyze the situation or tell the recruiter what to offer. One of the team's principles, Castellanos says, is to own what your AI does, which is easier to honor when a human is the only party capable of concluding anything.
The old version had a recruiter waiting a day or two for a compensation partner, then negotiating with someone who lacked the candidate context. Compadre didn't make that meeting faster. It deleted the meeting, and the handoff between two functions went with it.
What You Get Back
The reason this sits on a People leader's desk rather than IT's is what it does to jobs. A prompt library makes a practitioner faster at work they were already doing. A skill takes the reformatting and the manual calibration that sat between the judgment moments and hands the judgment moments back.
Every artifact Wilkins' hiring workflow produces is a draft rather than a decision, and two weeks of fragmented effort across five tools became a few hours of human review.
The cost is integration. A skill starved of data can’t do much that a prompt can’t. Many platforms that pitched integrated AI systems in 2024 delivered a dashboard with a chatbot bolted on. So the real choice isn’t about procurement. It’s about what you’re actually building.
Wilkins would back the practitioners who think in workflows over the ones who write clever prompts, and that's a hiring and development call.
