For a century, measuring a skill meant one thing: a test
Ask how to measure a skill and the reflex answer is always the same — set an exam, score it, done. Testing has been the default for so long that we treat it as the definition of measurement rather than one method among several. That reflex is now breaking, and not because tests are bad. It's because the best way to measure a skill is no longer a single method at a single moment. It's a layered, continuous read — one that increasingly reaches into the work people already do.
To see why, it helps to lay the methods side by side. Each measures something real. Each misses something real. The art is knowing which, and combining them.
The measurement spectrum
Self-report and the résumé. The cheapest signal: ask, and they tell you. Instant and broad, but it's a claim, not a measurement — and generative AI has made a flawless, keyword-perfect claim free to manufacture. Useful for discovery. Worthless as proof.
Credentials. A durable record that someone once cleared a bar. Real, but generic and frozen — it says what was true on exam day, not what's true now, and rarely at the resolution a specific role needs.
Tests and assessments. A controlled, comparable sample. This is a genuine measurement — everyone faces the same items, scored the same way — and it's the right tool when you need comparability at volume. But it's a snapshot of performance in an artificial setting, and it struggles with applied, contextual, judgement-heavy skill.
Adaptive interviews. A conversation that probes, follows up and adjusts difficulty in real time reaches depth a fixed test can't — and, done well, anchors every score to the evidence behind it with a confidence bound. The trade-off is that it's bounded in time and costs more attention per person.
Work evidence. The richest source of all: the deals, the code, the calls, the tickets people produce every day. It's real, contextual and continuous. But it's noisy — activity is not capability — and turning it into a defensible level takes interpretation, corroboration and consent.
"The best way to measure" is not a method — it's a fit
The mistake is to ask which method is best, as if one wins. None does. The best measurement comes from two moves. First, match the method to the skill: some skills prove out in a single episode and suit a test or a focused interview; others only prove out over a whole arc of work and demand evidence from the flow. Second, layer the methods so each covers the others' blind spots. A calibrated test establishes a comparable baseline; an adaptive interview adds depth and evidence; work evidence keeps it current. Any one alone is partial. Together they triangulate.
The frontier: measuring in the flow of work
The direction the whole field is moving is clear: away from pulling people out to be tested, and toward reading skill from the work they already do. This is measurement in the flow of work, and its appeal is simple. It's current — the read updates as the work happens, instead of freezing on exam day. It's real — it observes the job itself, not a proxy for it. And it's low-friction — nobody schedules a test; the evidence is a by-product of doing the work.
A test measures how someone performs when they know they're being measured. The flow of work measures how they perform when they're just doing the job. For most skills that actually matter, the second is the truer read.
Why it's hard — and what makes it honest
If in-flow measurement were easy, it would already be everywhere. It isn't, because raw work is the noisiest signal of the lot, and turning it into a trustworthy level takes three things most tools skip.
- An ontology to land on. You can't measure a skill from a work signal unless you've defined what the skill is, what its behaviours look like, and what level means. A signal maps to a behaviour, a behaviour to a skill, a skill to a level on a scale — none of which exists without the model underneath. In-flow measurement is impossible on a flat taxonomy; it needs the whole skills ontology.
- Corroboration, not one-shot. A single email, call or commit is a lead, not proof. Evidence becomes a level only when it's confirmed across multiple episodes and sources — and even then, outcome-linked skills carry lower confidence because attribution is genuinely hard.
- Consent and human review. Reading skill from work is one careful sentence away from surveillance. The difference is structural: people opt in, evidence is validated before it counts, and the close calls go to a human. Validated, not watched.
What the future actually looks like
Not a single method winning, but a loop. A controlled baseline — a well-built interview or assessment — establishes where someone stands, precisely and comparably. Then the flow of work keeps that read alive, refreshing it as evidence accumulates and flagging when it goes stale. Measurement stops being an event you schedule and becomes a property of the work, continuously corroborated and always current. The test doesn't disappear. It becomes the first reading in a living measurement, not the only one.
Key takeaways
- Testing is one method of measuring a skill, not the definition of it — self-report, credentials, tests, interviews and work evidence each measure something and miss something.
- The best measurement isn't a single method: match the method to the skill, then layer methods so each covers the others' blind spots.
- The frontier is measurement in the flow of work — current, real and low-friction, because the evidence is a by-product of doing the job.
- It only works on a skills ontology (signal → behaviour → skill → level), with corroboration and consent — the difference between validated evidence and surveillance.
- The future is a loop: a controlled baseline kept continuously current by the flow of work — measurement as a property of work, not a scheduled event.