Replicating Spotify's Claude Code plugin cost more tokens, not less

An experiment replicating Spotify's token-saving "shunt" plugin for Claude Code resulted in higher token usage and slower performance, contradicting the company's reported 90% savings. The test found the plugin's hooks were often ignored and delegated writing tasks required costly verification. The user rebuilt Spotify's two AI modes using local Qwen3-Coder and Claude Haiku as cheaper worker models. However, Claude Opus 5.5 already read files selectively via grep and sed, bypassing the plugin's hooks. Delegated test-writing consumed 23-34% more tokens and took 7-10 times longer, as Opus had to review and fix flawed drafts. Spotify's savings likely apply to workloads involving large, whole-file reads, which the test didn't trigger. The author concludes the plugin's value depends on specific usage patterns, and users should check their own token metrics before adopting it.