The problem
HelioGrid's Python services touch SCADA, billing, and a 20-year-old Oracle DB. Every junior engineer hit the same five gotchas in their first month. Senior engineers spent 4–6 hours/week answering the same questions.
Sprint 1 — Identify the recurring questions
We pulled six months of internal Slack #help-engineering, clustered the questions, and ranked them by frequency × time-to-resolve. The top 40 covered ~85% of inbound.
Sprint 2-4 — Build 40 Skills
Each Skill lives in .claude/skills/{name}/ with:
SKILL.md— the prompt and the auto-activation descriptionreference/— runbooks the model can readscripts/— small helpers (e.g.tail-scada-events.sh)
Skills include oracle-debug, scada-replay, billing-recompute, grafana-dashboard-create, pager-rotate-schedule, etc. The descriptions are written to auto-activate when the conversation matches.
Sprint 5-8 — Track usage
A PostToolUse hook logs which Skills auto-loaded and whether the conversation ended in a resolution. Top Skills get attention; rarely-used ones get reviewed.
What stuck
Skills are how a team levels up its juniors without burning senior time. The auto-activation description is half the value — write it like a help-desk ticket title.