Agent skills for benchmark reproduction with Claude Code and Codex.
| Configuration | Claude Code | Codex |
|---|---|---|
| Reproduce (Skill-guided) | Reproduce | Reproduce |
| Builder + Reproduce | Builder + Reproduce | Builder + Reproduce |
Reproduce follows the declared workload through preparation, preflight, execution, repair, verification, and reporting. Builder + Reproduce first constructs a source-grounded benchmark contract and assesses reference quality against the MLCommons ontology, then hands that contract to Reproduce. Reference-quality scores and reproduction agreement are separate assessments.
Choose the packages for your agent. Install reproduce for Skill-guided.
For Builder + Reproduce, install both builder-reproduce and reproduce from
the same agent directory. The builder package contains the ontology reference,
rubric, templates, and scoring scripts; the reproduction procedure is shared
with the standalone Reproduce configuration.
- Claude Code: place the selected folders under your project's
.claude/skills/. - Codex: place the selected folders under your project's
.agents/skills/.
Invoke reproduce for reproduction alone. For Builder + Reproduce, invoke
builder-reproduce, request the ontology assessment and score, and continue
with reproduce in the same session. Builder scoring requires Python 3 and
PyYAML (python -m pip install pyyaml).