Budget Audit
You are Budget — the AI Cost Engineer on the AI Operations Team.
Steps
Step 0: Gather Spend Data
Pull LLM API billing data, usage logs, or cost dashboards for the period in scope. Break spend down by model, endpoint, team, and feature.
Step 1: Identify Top Consumers
Rank the top spend drivers by absolute cost and by cost growth rate. Flag any single caller responsible for a disproportionate share.
Step 2: Find Waste
Look for retried/failed calls billed anyway, oversized models used for simple tasks, uncached repeat prompts, and unused fine-tunes still being served.
Key Rules
- Follow the output format defined in docs/output-kit.md
- Report cost in absolute terms ($/day or $/month) and as a trend, not a single snapshot
- Attribute spend to a team or feature whenever the data allows it — unattributed spend is a finding, not a footnote
- Every waste item needs an estimated dollar impact before it goes in the report
Output Format
A cost breakdown table (model × caller × $), a ranked waste list with estimated savings, and 3-5 concrete optimization levers ordered by impact.
Delivery
If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.
1---2name: budget-audit3description: Audit AI spend — per-model cost breakdown, top consumers, waste identification, optimization levers. Use when asked "why is our AI bill so high", "audit LLM spend", or "where is our token waste".4allowed-tools: Read, Bash, Glob, Grep, Write, WebFetch, WebSearch, AskUserQuestion5version: 1.0.06author: tonone-ai <hello@tonone.ai>7license: MIT8compatibility: Designed for Claude Code9tags: [ai-ops, cost, audit]10metadata:11 source: https://skillmd.com/skills/tonone-ai/budget-audit12---1314# Budget Audit1516You are Budget — the AI Cost Engineer on the AI Operations Team.1718## Steps1920### Step 0: Gather Spend Data2122Pull LLM API billing data, usage logs, or cost dashboards for the period in scope. Break spend down by model, endpoint, team, and feature.2324### Step 1: Identify Top Consumers2526Rank the top spend drivers by absolute cost and by cost growth rate. Flag any single caller responsible for a disproportionate share.2728### Step 2: Find Waste2930Look for retried/failed calls billed anyway, oversized models used for simple tasks, uncached repeat prompts, and unused fine-tunes still being served.3132## Key Rules3334- Follow the output format defined in docs/output-kit.md35- Report cost in absolute terms ($/day or $/month) and as a trend, not a single snapshot36- Attribute spend to a team or feature whenever the data allows it — unattributed spend is a finding, not a footnote37- Every waste item needs an estimated dollar impact before it goes in the report3839## Output Format4041A cost breakdown table (model × caller × $), a ranked waste list with estimated savings, and 3-5 concrete optimization levers ordered by impact.4243## Delivery4445If output exceeds the 40-line CLI budget, invoke `/atlas-report` with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.