Calls + amplification
Providers, models, endpoints, retries, agent loops, fan-out, regenerations, fallback cascades, and duplicate work.
Free open-source agent skill
Review production AI costs from the code outward.
Give Codex or Claude a disciplined method for locating AI call paths, verifying cost controls, separating evidence from inference, and producing the instrumentation plan needed for an economic decision.
03 / Download the skill
Evidence first. Conclusions second.
The review maps providers, models, production call sites, wrappers, retries, loops, context assembly, caching, output controls, fallbacks, quality safeguards, and usage attribution.
- Exact file-and-line evidence
- Observed, inferred, and runtime-data conclusions separated
- Read-only review unless implementation is requested
- Codex and Claude-compatible package
Download the skill MIT licensed · v1.0
What it reviews
The cost mechanisms hidden across a call path.
Source review is most useful when it connects the model call to the workload, retries, context, quality controls, and downstream outcome—not when it produces a list of SDK names.
Tokens + controls
Repeated prompts and tools, retrieved context, conversation history, output limits, caching, batching, and model configuration.
Measurement + quality
Usage capture, workload and customer attribution, latency, outcomes, evaluation sets, quality floors, rollout, fallback, and rollback.
Example prompt
$production-ai-cost-review Audit this repository for production AI cost controls and measurement gaps. Do not change any files.
What a source review cannot prove
Code alone cannot establish live traffic, provider charges, cache-hit rates, workload mix, quality, customer outcomes, verified savings, or migration payback. The skill identifies what is observable and specifies the runtime evidence needed next.