Three hours of prompt training gated the 28% PR gain in a 70-engineer rollout

Weeks one and two show the clearest signal in the design. Access went out with no guidance at all, deliberately, so the team could watch organic adoption ^[5]. What the instrumentation caught was divergence. The author reports the prompting skill gap widened, and that the widening was largest in the four weeks before training ran ^[19].

The arithmetic behind that spread is stark. A senior with strong prompting at roughly 40% velocity improvement, a junior with weak prompting at roughly 5%, is an eight-to-one difference in outcome from the same tooling on the same codebase ^[19][2]. The author is explicit that the results came out more nuanced than the vendor benchmarks suggest ^[22].

So price the enablement. Two 90-minute sessions per squad is three hours of contact time per engineer, which across 70 engineers is about 210 engineer-hours if everyone attends both ^[6][1]. Add the weekly retros running through months two and three ^[7]. Add a second round of teaching, because in month one review times fell for the wrong reason: reviewers were approving faster rather than reading better, and correcting that took explicit re-training ^[18]. That cost lives outside the seat invoice.

For the 28% to transfer, several things have to be true of your team ^[9]. You need your own six-month baseline out of GitHub and Jira, or you are comparing against a memory ^[4]. You need the seniority mix that produced the gain, since engineers with two to five years of experience saw the largest boilerplate wins while seniors who had already internalised those patterns saw less ^[13]. You need coverage headroom, because the biggest reported win was test writing and coverage moved 13 points ^[11][12]. And you need the review culture rebuilt, which is the author's own headline finding: obvious errors fall, while plausible-but-wrong logic errors hold flat or rise if review does not adapt ^[17].

The coverage-versus-defects gap is the number I would interrogate first. Thirteen points of new tests bought a 7% improvement in escape rate ^[11][10], and two logged incidents ran the other way when engineers over-trusted generated domain calculations that were syntactically fine and semantically wrong ^[16]. Coverage is a lines-executed metric, a separate question from whether invariants are asserted. That reading is mine, not the study's.

One more limit on what this data can settle. Claude Code was primary, Copilot ran on a single squad, and Cursor on three frontend engineers, about 4% of the team ^[3][4]. There is no control group held still, participation was voluntary, and the two engineers who opted out had both opted back in by week six, which is the closest thing to a unanimous verdict in the dataset ^[8].

In my context this reorders the plan more than the budget: book the training dates first, then switch on access, because the pre-training window is where the reported damage shows up ^[19].