Token Spend Auditor

Token Spend Auditor audits an AI agent's prompts, system instructions, and context setup, then hands back a structured report showing exactly where tokens are being wasted and how to fix it.

Most agents quietly bleed money through context bloat — redundant instructions restated in different words, conversation history that never gets trimmed, and system prompts that were never designed to be run thousands of times a day. This skill finds those leaks and fixes them with a fixed, repeatable six-pass process: context hygiene, prompt compression, context architecture, caching opportunities, budget guards, and a final cost report.

The output isn't a vague "here's a shorter version." You get a before/after token estimate, a ranked list of savings opportunities with the reasoning behind each one, a lean production-ready rewrite that keeps every original constraint intact, and concrete budget guards — token caps, escalation rules, and loop limits — to stop runaway agent spend before it happens.

It works standalone with no code execution, tokenizer library, or API access required, so it runs the same way in any SKILL.md-compatible environment, including restricted or no-code-interpreter setups.

Best for: AI agent builders, automation agencies, SaaS teams running high-volume agents, and anyone preparing a prompt or agent workflow for production who wants the savings to be measurable instead of guessed.

Known limitation: token estimates use a disclosed ~4-characters-per-token heuristic and may vary slightly from a specific model's real tokenizer. It does not pull live API billing data — if you have real usage numbers, feed them in and the skill will use those instead.