Sample App Benchmark

Use this skill when the question is whether a new system, workflow CLI, or instruction layer actually helps on real app work.

This skill standardizes the benchmark shape:

• one clean source repo
• main as the committed baseline
• one committed challenger branch
• one top-level Codex session per arm
• one isolated benchmark CODEX_HOME per arm using the sample_app_eval profile
• one normalized post-run verifier for both arms
• no benchmark residue in the source repo

Use When

• benchmarking a new agent aid on a fresh sample app
• comparing docs-first guidance vs a challenger branch
• testing whether a workflow CLI improves discovery on real implementation work
• creating reusable benchmark evidence for multiple systems with the same method