Lean body that assumes Claude's competence — it does not explain Docker, git, or what a benchmark is, and every section earns its tokens with executable commands and non-obvious wiring notes.

Provides fully executable `pier run` commands with exact flags, OpenRouter slugs, env vars, and a copy-paste-ready single-task smoke test rather than pseudocode.

Sequences state-check → setup → wiring → smoke test → subset → full → output, with an explicit pass-criteria checkpoint before the costly full run and a failure-mode table for error recovery.

No external bundle files exist, but the self-contained ~90-line body is organized into well-labeled sections with clear navigation and no nested references, fitting the simple-skill guidance.

Lists multiple concrete actions — scoring models on the DeepSWE benchmark, single-task/subset/full 113-task runs, and leaderboard submission — rather than vague language.

Explicitly states what the skill does (score AI models on DeepSWE via OpenRouter) and when to use it via an explicit 'Use when...' clause with concrete triggers.

Quotes natural phrasings a user would actually say ('run DeepSWE', 'benchmark this model on DeepSWE', 'score model X on the coding benchmark'), giving good trigger coverage.

A clearly scoped niche — coding-agent eval on the DeepSWE benchmark via OpenRouter — with distinct triggers unlikely to fire for unrelated skills.