context budget intelligence evaluator

The problem

Total file size is a poor metric for AI skill performance. The real issue is whether the worst-case simultaneous context load leaves enough headroom for the model to reason effectively without hitting token limits or degrading output quality.

What it does

• Inventories token counts for frontmatter, main body, and reference files individually.
• Maps specific trigger paths to determine exactly which files load for different tasks.
• Calculates peak context footprint based on the single most demanding task path.
• Benchmarks footprint against 8K, 32K, and 128K ceilings to identify reasoning-quality risks.
• Audits "lightweight" claims against actual measured data to ensure marketplace honesty.

Why this beats prompting it yourself

Most LLMs default to summing the entire directory, which overestimates context usage for skills using progressive disclosure. This tool models actual routing logic to find the real peak load and identifies specific structural bloat that simple token counters miss.

Use cases

• Audit a complex skill package before marketplace submission.
• Test if a high-performance prompt can be ported to 8K-context edge models.
• Identify which specific reference files are causing reasoning degradation.
• Verify that a "minimalist" skill isn't secretly bloating the context window.

Known limitations

Provides approximate token counts based on character heuristics. It identifies context-loading risk but cannot confirm actual reasoning degradation without user-provided side-by-side outputs.