Crossreview — AI skill for Claude Code

crossreview

An agent skill for **cross-review**: the agent you are working with sends one review brief to several other console code agents at once, Claude Code, Codex, Coddy, Cursor Agent, Devin, OpenCode and others, each on the model you picked, collects every answer, checks each finding against the code and decides what is worth fixing. One model reviewing its own work misses what a different model from a different vendor notices; this skill makes asking four of them as cheap as asking one.

It is a portable skill: install it in Claude Code, Codex, Cursor, OpenCode, Coddy, Devin, Kimi or any agent that reads `SKILL.md`, on Linux, macOS or Windows.

How a review goes

                 brief.md (the diff, the intent, the answer format)
                            |
     +-------------+--------+------+-------------+--------------+
     v             v               v             v              v
  claude -p    codex exec     agent -p       coddy -i      devin -p      ...  each alone, read-only,
  (sonnet)     (gpt-5.6-sol)  (auto)         (devin/swe-2) (swe-2-high)       blind to the others
     |             |               |             |              |
     +-------------+--------+------+-------------+--------------+
                            v
          the orchestrator (your agent): collects every answer,
          verifies each finding in the code, decides: fix / not worth fixing / rejected / open

You invoke the skill, and the agent you talk to asks what it must (which reviewers, what the change is for), writes the brief and hands the rest to a subagent: Coddy's built-in `crossreview` coordinator, a `general-purpose` agent in Claude Code, a task in OpenCode. The subagent runs the reviewers, collects every answer, checks the findings and decides, and its report comes back as the answer, so none of that work fills your conversation. An agent without subagents does the same steps itself.

Three rules hold for every run:

• **Reviewers are blind t