Choose skills vs subagents by repeat-frequency: skills amortize, subagents isolate but reset every call

alexop.dev·low signal

Skills load markdown into the live context (a dedup cache loads each at most once per session), while subagents get an isolated window and reset on every invocation — for one multi-domain query, isolated subagents used ~9K tokens versus ~15K for the context-accumulating skill pattern. But subagents give zero amortization on repeated similar queries because each call starts cold, and near-match skill templates can severely pollute the main context. The rule of thumb: route one-off, contamination-prone work to subagents and frequently-repeated, stable instructions to skills.