Cost & Performance

Two complementary techniques for controlling Claude Code spend — one on the output side, one on the measurement side — plus the honest pushback on over-claimed retrieval tooling.

What you’ll find here

• Output compression — the caveman plugin (75% output-token reduction, paper-backed brevity-improves-accuracy claim)
• Production observability — Rezvani’s OpenTelemetry + Docker Compose stack, the 8 metrics that matter, TRACEPARENT subprocess propagation
• The honest reframe on retrieval — Graperoot’s 178× headline was satirical; real savings are 50–85%, plus community concerns about vendor trust

Featured in this category

1. Cut Claude Code’s Output Tokens by 75% — two commands, measurable savings
2. The New Claude Code Monitoring — what Rezvani’s 7-person team found in 2 weeks of telemetry

Combined effect

Used together, the caveman plugin (output-side) and the OpenTelemetry stack (measurement-side) typically produce 50-75% token reduction with visible per-prompt cost attribution. The Graperoot article is worth reading to recalibrate expectations around inflated retrieval-tool marketing.

---

Table of contents

• Cut Claude Code's output tokens by 75%
• The New Claude Code Monitoring
• I reduced my token usage by 178x — the honest reframe