Skills-Based Hiring: What It Is and How to Actually Do It

Skills-based hiring replaces the question *"has this person held a similar job?"* with *"can this person do this work?"* Those sound similar and are not. The first is a proxy; the second is the thing you actually want to know.

The mechanics are straightforward. Instead of a job description asking for "5+ years and a relevant degree," you define the role as a set of skills, each with a required proficiency level. Instead of screening CVs, you assess those skills directly. Instead of ranking on pedigree, you rank on evidence.

A CV is a compressed, self-authored, unverified claim. Three consequences follow, and they compound.

It measures writing, not working. Someone who learned product management inside an engineering role has no slot on a CV to say so. The format has one line per job — title, company, dates — and everything they actually learned lives outside it. They lose to a worse candidate who writes a better document.

It forces employers into proxies. There is no degree for "AI Builder" and no career ladder that leads to it. So the job ad asks for years and keywords, which exclude the person with three shipped side projects and include the person who wrote "AI" in status reports for five years.

It hides the disagreement. The candidate is thinking about their ability. The employer is reading a document about that ability, filtered through the candidate's self-marketing skill and the recruiter's pattern-matching. There is no shared artifact either side can point at.

The mechanical change is small and the mental change is large: express seniority as a required proficiency level on a specific skill, not as years.

Prompt Engineering — Level 3 (required) instead of 5+ years experience with LLMs.

For this to work the levels need behavioural definitions. "Advanced" means nothing. "Decomposes unreliable tasks into verifiable steps and builds an evaluation set before changing anything" means something, can be assessed, and can be argued about productively with your hiring manager.

Practical constraint: keep it to four to six skills. Specs with twelve required skills are wish lists, and they reintroduce the filtering problem you were trying to remove.

Long take-homes are extractive and they select for candidates with free weekends — which is its own bias. The fix is not length, it is decision density.

A well-designed 60–90 minute challenge presents a messy, realistic situation with a genuine trade-off in it, and scores the reasoning rather than the volume of output. Our refund-agent challenge asks for a design, not code, and the highest-weighted criterion is whether the candidate says what they refuse to automate.

Three rules that make assessments fair: publish the rubric before the candidate starts, cap the time, and make the resulting proof portable so the candidate does it once and it counts everywhere.

This is the first objection everyone raises, and it is usually the wrong frame. For most modern roles, using AI well *is* part of the job. Banning it assesses a working environment nobody has.

The productive design allows AI use openly and targets criteria an assistant will not supply: what you chose not to build, where you put the approval gate, which evidence you weighed and which you discounted, and how you would know your fix worked. Ask an assistant to "prioritise this roadmap" and it produces a balanced answer that cuts nothing — a pattern that is easy to score and caps at the middle of the scale.

Add a short reflection question about the next step, and the gap between judgment and generated output becomes visible quickly.

It can, and it is not automatic. Removing names, schools, and employers from the shortlist removes some of the strongest proxies for class and background. Publishing the rubric makes bias in the instrument itself visible and arguable, which a CV screen never is.

But an assessment can encode bias as easily as a CV screen. Guardrails that matter: exclude language fluency and formatting from the rubric where they are not the skill being measured, sample-audit scores with human reviewers, and publish outcome distributions where you lawfully can. An unexamined algorithm is just a faster version of the problem.

Pick one role — ideally one you have struggled to fill, because that is where the proxy is most obviously failing. Define it as four to six skills with required levels. Send every applicant the same challenge link. Rank on the results, blind.

Then check one number: how many people in your top ten would your old CV screen have rejected? That number is the entire business case, and it is usually between three and six.