Getting good at Claude: A research-backed curriculum | Claude

Detailed Analysis

Anthropic's AI Fluency Index, initially launched to study how users develop proficiency with Claude Chat, has expanded to encompass Claude Code and Claude Cowork, drawing on data from over 50,000 conversations across 11 behavioral indicators. The research reveals a foundational distinction in how fluency develops: some skills emerge organically through repeated exposure, while others must be deliberately and repeatedly taught. This bifurcation—between what Anthropic terms the "Description spectrum" and "Discernment"—forms the backbone of a curriculum model the company is offering to organizations seeking to train their workforces in effective AI collaboration. The expansion beyond Chat to agentic surfaces reflects the maturation of Claude's product ecosystem and the growing organizational stakes of AI adoption.

A central finding of the Index is that each Claude surface has what Anthropic calls a "signature move"—a gateway behavior whose presence most reliably predicts broader fluency. In Chat, that behavior is iteration: users who refine their outputs through follow-up turns demonstrate stronger performance across every other measured dimension, while single-turn users show virtually no critical evaluation at all. In the agentic surfaces of Claude Code and Claude Cowork, the signature move shifts to goal clarity—explicitly stating objectives before Claude begins working. The distinction is not incidental. Agentic workflows, where Claude takes sequential autonomous actions rather than generating a single response, are particularly sensitive to vague initial inputs; downstream errors compound in ways they do not in a back-and-forth chat interface. This means that onboarding curricula must be surface-specific, not generic.

The Description spectrum—the range of behaviors through which users shape Claude's outputs—shows encouraging self-sufficiency. Users with longer tenure naturally gravitate toward more sophisticated shaping behaviors: providing examples, setting interaction styles, communicating tone expectations. Critically, this development appears non-linear and does not require intensive instruction; exposure itself drives growth along this dimension. The implication for training designers is one of prioritization. If curriculum time is scarce, the Description spectrum can largely be seeded through practice and exploration rather than formal instruction. The spectrum is organized by "durability," from in-the-moment adjustments like file uploads and iteration, to persistent configuration tools like Projects, CLAUDE.md files, and scheduled workflows—a hierarchy that also maps roughly onto sophistication of use.

Discernment—the evaluative behaviors surrounding what Claude produces—is identified as the most critical and most resistant area of development. Unlike Description skills, Discernment does not emerge with tenure, does not transfer from feature familiarity, and is actively undermined by a shift toward observational verification. As Claude increasingly produces outputs users can directly inspect—diffs, reports, compiled code—the temptation to substitute looking for questioning grows. Yet observational verification is structurally blind to certain categories of error: false assumptions, missing context, and confident-sounding but factually wrong claims. Anthropic also flags a structural workforce concern: as tasks historically performed by early-career employees are automated by AI, the institutional knowledge of "what good looks like" risks being lost rather than transferred, making intentional Discernment training not merely pedagogically useful but organizationally necessary.

The curriculum model Anthropic proposes—teach the signature move first, advance along the Description spectrum, revisit Discernment at every step—reflects a broader trend in enterprise AI adoption: the recognition that deploying AI tools is insufficient without parallel investment in human capacity to use and evaluate them effectively. The framing of Discernment as something that "has to be taught" rather than something that develops naturally represents a pointed organizational warning at a moment when many companies are racing to maximize AI-driven efficiency. The Index positions Anthropic not only as an AI developer but as an active participant in shaping how organizations build long-term human-AI collaboration literacy, a role that aligns with growing industry and regulatory attention to AI literacy as a distinct and teachable competency.