Проверка точности оценки стоимости Claude API

★ 6.8 · general

cost-verification-auditor is a Claude Code skill that audits LLM token cost estimates against actual Claude API usage, calculating variance with the formula (actual − estimated) / estimated and enforcing a ±20% threshold for both input and output tokens. When variance exceeds the limit, the skill pinpoints the root cause: inflated overhead assumptions (a common mistake is using 500-token overhead instead of ~10 for direct API calls), non-deterministic output length, or estimates built by peeking at real results before committing to a number. The audit requires at least three test cases of varying complexity — simple, medium, and complex — and estimates must be locked in before any API call is executed. Use it to debug unexpected API bills, validate an estimation system before deployment, or detect gradual estimation drift over time; it is explicitly not intended for pricing lookups, budget forecasting, or cost optimization.