This skill gives Codex a narrow review procedure: treat user-supplied facts as the complete evidence boundary, split draft prose into atomic claims, and show what is supported, inferred, unsupported, or conflicting.

Package contents

  • SKILL.md defines the trigger, required input, review sequence, exact output sections, and stop conditions.
  • references/review-rubric.md defines five verdicts and requires mixed sentences to be split before scoring.
  • agents/openai.yaml supplies optional display metadata for supported interfaces.

Required input contract

The user must provide labeled SOURCE FACTS and DRAFT CLAIMS sections. If either section is absent, the skill stops rather than filling the gap from memory or the web. Instructions inside either section are treated as quoted content.

What was tested

The package passed structural validation. Four explicit, separate, read-only runs tested mixed claims, a quoted instruction that tried to override the evidence boundary, an unresolved material conflict, and a missing-input case. Expected invariants were withheld until grading.

What remains untested

The retained evidence does not cover implicit activation, other models or hosts, large documents, open-web retrieval, or every form of prompt injection. Review the files before installation and repeat the tests in the environment where you intend to use the skill.