Add llm-cost-optimization skill

[vikast908]

What

A new llm-cost-optimization skill on cutting runtime LLM token and dollar cost without lowering output quality.

Why

The repo has context-engineering (the coding agent's own session) and performance-optimization (web vitals and latency), but nothing on the API cost of the model calls a product makes at runtime. That cost is usually dominated by avoidable waste: a static system prompt resent uncached on every call, full history replayed each turn, whole documents stuffed into context, unbounded output, and a frontier model doing cheap work.

What's in it

• Measuring cost per request before optimizing
• Caching the static prefix (system prompt, tool schemas, examples)
• Windowing or summarizing history instead of full replay
• Narrow retrieval instead of stuffing context
• Capping and structuring output
• Routing cheap, high-volume tasks to a smaller model
• A "don't trade away quality" guard, with a Verification checklist

Conventions followed

• Standard anatomy (Overview, When to Use, Process, Common Rationalizations, Red Flags, Verification).
• Frontmatter is name + description only; description leads with what it does, then Use when triggers.
• Single SKILL.md, no supporting files.
• References context-engineering and performance-optimization rather than duplicating them.

Related: this pairs with my other PRs adding reliable-agent-loops (#285) and evaluating-llm-output (#286). Happy to adjust naming, scope, or split anything.

🤖 Generated with Claude Code