How we measure AI fluency.

What "fluency" means here.

"AI Fluency is the ability to work effectively, efficiently, ethically, and safely within emerging modalities of human–AI interaction."

We adopt this definition. It captures what other framings miss: that fluency is not "uses AI a lot" or "knows the tools." It's a craft: knowing what to ask for, how to hand it off, when to push back, what to double-check, and whether the work actually ships.

But a definition isn't a measurement. To turn fluency into something you can see and improve, we built a five-dimension rubric scored from your actual conversations.

The 5 D's.

Each dimension sits at a different point in the human–AI workflow. They are scored independently, on equal footing, against the same 0.0–10.0 scale to one decimal place.

We re-score every two weeks so you can see your trajectory, not just a snapshot. Decimals matter: the difference between a 7.2 and a 7.6 is the difference between entering a band and solidly in it.

Where the framework comes from.

The first four D's, Description, Delegation, Discernment, Diligence, build on the foundational AI fluency work of Prof. Rick Dakan (Ringling College of Art and Design) and Prof. Joseph Feller (University College Cork), in collaboration with Anthropic. It is the most widely cited academic framework in the field.

We add a fifth dimension, Delivery, because the academic framework deliberately ends at responsibility, and in our data the gap between insight and shipped work is where fluency either compounds or evaporates. A score that doesn't account for whether the work actually crosses the line into real output overstates fluency for users who never ship.

All five dimensions are scored at equal weight. There is no "core four plus extension" framing inside the product, in the analyzer, or in your profile. Delivery is graded against the same rubric structure and contributes to your overall score on the same terms as the other four.

The artifact-discernment gap.

In a study of 9,830 conversations published by Anthropic in February 2026, users who produced polished outputs, code, documents, apps, were measurably less likely to question the model's reasoning, identify missing context, or fact-check the result.

Lee et al. (CHI 2025) found the same pattern at the user level: higher confidence in AI predicts less critical thinking, and verification effort drops the moment AI output starts looking finished. METR's July 2025 study of experienced developers found that AI-assisted work shipped more regressions despite users feeling faster.

This is why Diligence in our rubric carries a Shipping Gate sub-metric. Discernment grades whether you push back during the conversation; Diligence's Shipping Gate separately grades whether you run an independent verification before the work leaves your hands. The data shows these decouple in real users. And the gap between them is exactly where the most expensive AI mistakes get made.

How a score is actually produced.

Scores come from the conversations you upload: never from a quiz, a self-rating, or a static benchmark.

1. You upload a corpus. A Claude or ChatGPT export ZIP, processed locally in your browser. The conversations leave your machine only for scoring; the raw text is never permanently stored.
2. We sample biweekly. Conversations are bucketed into two-week periods over your most recent 90 days. Within each bucket we sample for breadth, not just recency, so a heavy user week doesn't dominate the score.
3. We grade against a fixed rubric. Each dimension is graded against a public, versioned rubric. The grader is told to score on observable behaviors, repeatable patterns, not single moments, and explicitly forbidden from grading on how sophisticated a topic sounds or how smart you appear.
4. We publish a confidence flag for any biweekly period with fewer than three substantive conversations. Low-confidence periods are visually de-emphasized on your trajectory chart.
5. Re-grading is automatic. When we improve the rubric (we're on version 14), every existing profile is re-graded against the new prompt within 24 hours of release. Your score reflects the current rubric, not the rubric in effect when you first uploaded.

How we label confidence in a score.

Every score is published with one of four confidence rungs so a recruiter can tell at a glance how grounded the number is. The rung is computed from how many of your projects we could link back to the conversations we graded: never from a self-rating.

Self-Reported

Score derived from this person's own AI conversations only: no linked projects.

Verified

Score corroborated by at least one of this person's linked projects, with recent activity on the platform.

Strongly Verified

Score corroborated by multiple linked projects across more than one platform: the highest-confidence rung.

Evidence-Only

Score derived from this person's linked projects only: no conversation corpus on file.

Five things we don't claim to measure.

Honesty about limits is part of the rubric. Here is what we explicitly don't do:

Behaviors that happen outside the conversation.

Anthropic's own research (Feb 2026) found that more than half the behaviors in a complete fluency framework happen outside the chat: being honest with collaborators about AI's role, weighing the downstream consequences of sharing AI-generated output, deciding not to use AI for a task at all. We hit the same ceiling. We don't pretend otherwise.

Whether the task was a good use of AI.

We grade how you worked with the model. We don't grade whether the model should have been involved at all. That's a judgment call that requires context we don't have.

Implicit behaviors.

If you mentally fact-checked the model's claim against your own knowledge but didn't say so in the conversation, we can't see it. Our score will under-credit you.

Comparative ranking.

Tier labels (Novice → Native) are behavioral classifications, not rankings. We don't tell you you're in the top X% of users. We don't believe that framing helps anyone improve.

Anyone else's data.

Your score is graded only against the rubric, never against another user's profile.

Why your trajectory matters more than your score.

A single score, taken once, is a snapshot. Anthropic's own data shows the most fluent users are the ones whose scores moved. Specifically, the single strongest correlate of every other fluency behavior is iteration and refinement: building on previous exchanges instead of accepting the first answer. In Anthropic's sample, conversations that iterated were 5.6× more likely to involve users questioning the model's reasoning, and 4× more likely to identify missing context.

This is why your VibeWorks profile re-scores every two weeks, why we publish your trajectory alongside your score, and why our coaching loop nudges you toward iteration before anything else. The score tells you where you are. The trajectory tells you whether you're learning.

The research this rubric is built on.

Every dimension and every behavioral signal in the rubric is grounded in published work. The grounding is named in the prompt itself, and we attribute it here so anyone can check our work.

We change. Transparently.

The rubric is versioned. Every previous version is documented. When we change the rubric, every existing profile is automatically re-graded against the new version within 24 hours; you can see your score under each version on your trajectory chart.

Your conversations stay yours.

Three things you should know:

• We never store your full conversation text. The grader sees a stratified sample (capped at ~90,000 characters across 90 days) for the duration of the analysis run, and that sample is purged on a 90-day rolling window.
• We never name specific products, projects, companies, employers, clients, or collaborators in any output we generate from your conversations. The grader is explicitly forbidden from doing so.
• We never quote your conversations verbatim back to you, or to anyone else. Standout examples describe what the behavior was and why it demonstrates skill: never what the work was about.

For the full privacy and security picture, see /security.

The instrument we wish existed when we started.

We built VibeWorks because the existing answer to "how good are you with AI?" was a quiz, a self-rating, or a vibe. This rubric is measurement.

If you find a flaw in it, a behavior we under-credit, a citation we missed, a tier ladder that misclassifies your work, write to us at research@fluencity.ai. Every published version of this rubric so far has been improved by user pushback. The next one will be too.