SkillMark
SkillMark reports four dimensions (Designing, Supervising, Reviewing, Ownership), each built on one of the four competencies of Anthropic's AI Fluency framework (Delegation, Description, Discernment, Diligence) and the 24 behavioral indicators underneath them — verified by what a person actually did, not by what they said they did. No black boxes, no inflated precision.
The same methodology serves two uses: a hiring team deciding whether to move a candidate forward, and an organization assessing or re-assessing its own employees' (participants') AI-fluency readiness over time. Scoring, indicators and Signals are identical either way — only the review screen's wording (e.g. follow-up questions framed for an interview versus a development conversation) adapts to which one a session is.
SkillMark's four dimensions map 1:1 onto Anthropic's AI Fluency competencies:
The specific behaviours we look for under each dimension — the underlying scoring rubric and the 24 behavioral indicators — come from Anthropic's published framework. See AI Fluency: Framework and Foundations and the 4 Ds of AI Fluency — Behavioral Indicators on Anthropic Academy.
SkillMark is not affiliated with or endorsed by Anthropic. We credit their framework because it is the common vocabulary — a candidate may already hold Anthropic's own AI Fluency certificate, and can bring it here to be verified. Everything above the indicators — the Probes that elicit evidence, the Signals computed from a session, and how a verdict is reached — is SkillMark's own model, described below.
Our pitch to a candidate: measured against Anthropic's 24 AI-fluency indicators — verified by what you did, not just what you said.
Delegation
Was work split sensibly between person and AI, with the person staying involved?
Deciding what to hand to AI and what to keep, and setting the work up so that split makes sense.
Indicators: Clarifying the goal before asking for help, consulting the AI on approach, recognising when a task isn't right for AI, staying involved rather than just accepting, choosing the right tool, adjusting who does what as work evolves.
Signals: SkillMark Signal: Delegation Fit (was the task partitioned into sensible parts before delegating; AI share is context, not a score).
Direction (Description)
How clearly were the goal, format, constraints and examples given — including tests as spec?
Steering the AI while it works: clear goals, constraints and context, and correcting course.
Indicators: Specifying format and structure, breaking complex tasks into steps, building on previous exchanges to refine, providing examples of what good looks like, setting boundaries on what the AI should or shouldn't do.
Signals: SkillMark Signals: Requirement Coverage (key requirements specified to the AI) and Hidden gems discovered (hidden facts surfaced by asking good questions).
Discernment
Was AI output checked, challenged and corrected — were traps and AI errors caught?
Checking what the AI produced: catching errors, questioning claims, fixing before accepting.
Indicators: Checking facts and claims that matter, noticing when the AI's reasoning doesn't hold up, recognising when the AI is missing context, assessing whether the collaboration is making progress.
Signals: SkillMark Signals: Traps and AI mistakes caught (planted traps avoided and AI errors challenged) and Loop Closure (problem threads that closed rather than stalled).
Diligence
Was the work shipped responsibly — safe sharing, honest about AI's role, tests green, responsibility taken for what was submitted?
Standing behind the result: handling information safely, being open about AI's part, finishing the job.
Indicators: Being mindful about what information is shared, honest about the AI's role in the work, taking responsibility for what is submitted, considering the consequences of putting the work into the world.
Signals: SkillMark Signal: Test Rigor (tests run, green at end, edge cases added). How much the AI wrote is never scored (Diligence: “owning the final result fully, regardless of AI’s role”).
A behaviour the task never gave an opportunity to show — for example, a timed assessment rarely surfaces who the intended audience of a document is — is never held against a candidate. That indicator reads as not applicable and is left out of the axis entirely, rather than scored as a failure. Anthropic's list also includes a few indicators that are genuinely rare to observe in a short, timed session; when the evidence for one is present, it can only raise the axis, never lower it.
A reviewer also sees “Fitness to role” — how closely a session's four-dimension profile matches the ideal profile for the role being assessed. It is a measure of fit against that ideal, not a verdict on the person, and it is never used to rank or shortlist anyone.
Anthropic's indicators describe general knowledge work. SkillMark also sees code, tests, a terminal and the final submission, and we instrument tasks with Probes — deliberate traps the AI gets subtly wrong, hidden context only revealed to a good question, and key requirements the candidate must both specify and meet — designed to elicit evidence a transcript alone would not show.
From those Probes plus the session, we compute named, versioned SkillMark Signals. Signals marked “scored” below feed an axis; the rest are descriptors, shown for context but never scored.
| Signal | Meaning | Scored? |
|---|---|---|
| Delegation Fit | Task partitioned into sensible parts before delegating (AI share is context, not a score) | Scored |
| Requirement Coverage | Key requirements specified to the AI, and met in the result | Scored |
| Hidden gems discovered | Hidden facts the person uncovered by asking a good enough question | Scored |
| Traps & AI mistakes caught | Deliberate traps avoided and AI mistakes spotted or challenged | Scored |
| Loop Closure | Problem threads that closed (discerned, refined, resolved) rather than stalled | Scored |
| Test Rigor | Tests run, green at end, edge cases added or tests enriched | Scored |
| AI share of final work | How much of the final work the AI wrote: context only, never scored | Descriptor |
| Delivery | Was the task solved — yes, partial, or no | Descriptor |
| Working Style | A profile of how someone works (Specifier / Explorer / Debugger / Verifier), not a rating | Descriptor |
How much of the work was delegated to AI is also a descriptor, not the Designing score itself — Designing measures the QUALITY of the delegation (goal clarified, right tool chosen, stayed involved), not the raw share of work handed over. Handing over more is neither good nor bad on its own.
Every weight behind an indicator or a Signal — how much each one contributes to its axis — is currently hand-set, not fitted against outcomes. None has been validated against whether the person went on to do the job well, and we do not publish the exact numbers: naming a precise weight per indicator invites optimizing for the number rather than the behaviour it stands in for, and the weighting is unvalidated (see above) — publishing false precision alongside an admittedly unvalidated model would be worse than not publishing it at all.
What we do publish is which indicators and Signals feed each axis (above), and the never-penalise rule for evidence the task never gave a chance to show. Saying so plainly is not a disclaimer; it is what makes the eventual claim of validation mean something.
The credibility anchor is transparency, not false precision.
Every scored session carries a scoring version. Sessions analysed before this indicator-based model shipped keep their earlier methodology and wording until they are explicitly re-scored — nothing is silently rewritten. A ranking, average or percentile never mixes sessions scored under different versions: it uses whichever version every session being compared shares, so a number is always comparable to the ones next to it.
None of the four axes references a vendor, a model, or a feature. How clearly you instruct, whether you verify what comes back, and whether you take responsibility for what you ship are properties of the working relationship, not of the tool on the other end. They read the same against a model released next year.
The assessments are built the same way. A candidate redlines a vendor contract, diagnoses a conversion drop, or traces a production bug — work the business already does. SkillMark measures whether someone can get that work done with AI, which is a slower-moving question than what the frontier is capable of this quarter.
Percentiles compare a person only against the same organization + same role cohort, scored at the same version — never a mix of an earlier and a later scoring methodology. A backend engineer's percentile is computed against other backend engineers in the same organization, not against all SkillMark users globally.
The cohort label and sample size (N) are always shown alongside any percentile. You will always know exactly what population a percentile is relative to.
Percentiles are shown only when the cohort contains 5 or more scored people at the same scoring version. Below that threshold a “benchmark building” indicator appears instead of a percentile.
Below 5 data points, percentile ranking is statistically unreliable — a rank of “80th percentile out of 3 people” carries almost no information. Rather than show a precise number that misleads, we acknowledge the limitation explicitly.
The vector itself is the portable signal during the low-data period. Each axis is interpretable on its own — no cohort comparison needed.
Sessions where the person used no AI prompts are excluded. Zero prompts means the analysis pass never had anything to read — there is no signal to score, and including it would dilute the result without adding information.
The session count shown on a credential reflects only the sessions that contributed.
The vector recomputes each time the person completes a new assessment and analysis finishes. The values shown on a credential are always the most recent computed ones, and the “as of” date reflects when that computation ran.
Scores are recency-anchored — they reflect the most recent valid session rather than an average across all of them. This preserves the improvement signal: someone who has grown their AI collaboration skills sees it reflected immediately rather than diluted by their earliest attempt.