Workflow quality is what the market means by AI fluency.
The gap between using AI and being effective with AI is now priced into the labour market. This page is about that gap, and what your llm-dx score measures within it.
Effective use is operational, not tool choice
The distinction that shows up in compensation data isn't about which models practitioners use or how often they use them. It's about how they use them. Structured sessions, disciplined context management, defined outputs, consistent evaluation: these are the operational differences between practitioners who extract compounding value from AI and those who extract inconsistent value at high token cost.
A 56% wage premium, doubled in twelve months
PwC's 2025 Global AI Jobs Barometer, drawn from nearly a billion job ads across six continents, found that workers with advanced AI skills earn a 56% wage premium over peers in identical roles. The year before, that premium was 25%. It more than doubled in twelve months.
The premium is attributed to AI fluency broadly. Workflow quality is the operational foundation of that fluency: not a peripheral concern but the primary variable separating effective from expensive usage.
Seven dimensions of operational discipline
The seven dimensions in this assessment map to the components of that operational discipline: project setup, knowledge quality, on-demand context, prompt quality, session discipline, efficiency, output discernment. None of them are about which model you're using.
A score isn't a credential. It's a diagnostic, showing you where your practice is solid and where it's costing you sessions, output quality, and time.
The construct is being measured at scale
In May 2026, Anthropic published its AI Fluency Report — a study of 11 observable fluency behaviors across 9,830 multi-turn conversations, built on the 4D AI Fluency Framework developed by Professors Rick Dakan and Joseph Feller in collaboration with Anthropic.
Two findings are directly relevant to what llm-dx measures. Practitioners who produce artifacts — code, documents, apps — are measurably less likely to challenge the model's reasoning or identify missing context. And augmentative use — treating AI as a thought partner rather than delegating work entirely — shows more than double the fluency behaviors of quick, transactional exchanges.
The first finding maps to Output Discernment, the dimension llm-dx practitioners most commonly score lowest on. The second is what the other six dimensions collectively enable.
The 4D Framework defines what fluent AI use looks like at population scale. llm-dx diagnoses where your workflow falls short of it — and gives you the correction path.
The same week Anthropic published this research, Gavin's own AI fluency report — generated by Anthropic's platform from 44 conversations — flagged one developing behavior out of eleven measured: challenging model reasoning. llm-dx maps that directly to the Output Discernment dimension (di3: examining the model's reasoning rather than accepting conclusions; di4: feeding errors back into skills and briefings rather than correcting only in-session). The platform identified the gap. The framework provided the specific correction.
Key takeaways
- Anthropic measured 11 fluency behaviors across 9,830 conversations using the 4D Framework (Dakan, Feller).
- Producing artifacts correlates with reduced critical reasoning (−3.1pp) and missed context (−5.2pp) — the gap llm-dx scores as Output Discernment.
- Augmentative use (AI as thought partner) shows >2× the fluency behaviors of quick, transactional exchanges.
- The 4D Framework defines fluent use at population scale; llm-dx diagnoses where your individual workflow falls short and gives you the correction path.
Where you stand
Find where your workflow stands. Fifteen minutes for the full assessment. Five for the Quick Check if you want your biggest gap fast.