When a global fintech with 2,400 engineers decided AI fluency was a table-stakes skill, they had a familiar problem: a handful of prompt experts on a central team, and everyone else copy-pasting from internal Slack threads.
Leadership set a goal that sounded impossible: every engineer, regardless of seniority or team, should be able to design, evaluate, and ship AI features to production standards within a single quarter.
Why Traditional Training Failed
Before LunaPrompts, they'd tried the usual playbook:
- Recorded courses with 12% completion rates
- Internal workshops that didn't scale past the first 200 engineers
- Certification badges that measured attendance, not ability
Nothing of what they tried produced measurable output changes. Prompts still regressed in production. PR reviews still flagged the same AI-adjacent mistakes. The central prompt team stayed the bottleneck.
The LunaPrompts Approach
The company rolled out LunaPrompts as their canonical evaluation layer for AI engineering skills. Three things made it click:
- Scored challenges over videos. Engineers practiced against real test cases, with automated grading, instead of passively watching.
- Team leaderboards. Weekly cohort rankings turned learning into a competitive loop the way code golf once did.
- Role-specific tracks. Backend, ML, and product engineers each got pathways that mapped to problems they actually hit.
Managers used challenge completion and accuracy as signals, not gates. That meant adoption stayed voluntary and engagement stayed high.
The Results (90 Days In)
- +68% accuracy on held-out production-representative prompt evaluations
- 3.2x faster ramp for engineers new to AI-adjacent projects
- 40% reduction in prompt-related PR review iterations
- 90-day full rollout, not the 6-month estimate initially budgeted
The central prompt team stopped being the bottleneck. They shifted to writing new challenges instead of reviewing every prompt in the company.
What Worked, What Didn't
Worked: Opt-in energy. Managers didn't force participation, but they made the leaderboard visible.
Worked: Role-specific tracks. A backend engineer debugging retrieval quality doesn't need the same challenges as a growth engineer drafting marketing copy.
Didn't work at first: Trying to skip the foundations. Even senior ICs benefited from re-running the basics.
Takeaway
"AI fluency" stops being a fuzzy goal the moment you measure it. The engineering org that can point to a scored, shared metric wins. The one still debating whether to run another lunch-and-learn doesn't.
LunaPrompts gave this team the metric. The rest was their own momentum.


