Introduction: Claude Code configuration totals 107 skills, 38 agents, and 15 commands, costing 9,857 tokens per session before any user input.

Results breakdown: Costs break down to 7,470 tokens for skills, 1,999 tokens for agents, and 388 tokens for commands, all summing to 9,857 tokens across 160 items.

Interpretation of impact: About 5% of a typical 200k-token window is consumed before any action, but the real issue is the high rent-to-trigger ratio that eats into context for other tasks.

Missed elements: Several advertised automatic components (delivery-gate, gateguard, safety-guard, ck, deep-research) may require extra wiring or are largely documentation, resulting in ongoing rent without clear utility.

Two costs of a skill: (1) load cost when triggered, which is visible and fair; (2) description cost that remains in the context window as permanent rent.

What the author did: Measured and deleted untriggered components over a month; the core script (_cc-tax_) is shared; a longer guide shows how to reduce a 284-skill setup to 104 and outlines criteria plus a cross-machine symlink-driven rules-library.

Heaviest contributors: Top costly descriptions include loop-design-check, token-budget-advisor, prompt-optimizer, intent-driven-development, and agent-architecture-audit, with a single skill (continuous-learning-v2) driving 56,453 tokens per trigger and signaling misalignment.

Overall takeaway: The cost is real but not catastrophic; the persistent rent-to-utility ratio is the core concern, and pruning unused skills can reclaim context space.

Call to action: Readers are invited to measure their own configurations; links to the script and GitHub repo Aliwers/cc-tax are shared, along with a paid longer version excluding non-functional elements; question posed: can anyone beat 9,857 tokens?

Script and method: A Python script (cc-tax) estimates token cost by parsing descriptions in skill, agent, and command metadata, converting characters to tokens, and aggregating costs per item and total.