Guide · Agent skill
How to use the Cross-Modal Review agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Cross-Modal Review skill (cross-modal-review) lets your AI agent send its work to a different AI model for an independent review. It is useful for developers who want a second opinion on significant code changes, and for teams making security-sensitive or bulk changes. After this guide you will know what the skill does, how to install it, how to trigger it with natural requests, and what it needs to run.
Cross-Modal Reviewgarrytan/gbrain · MIT Download the skill Quick start
- 1Download the ZIP. Download the cross-modal-review skill ZIP from this page.
- 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use, or into your project's .claude/skills/ folder for one project.
- 3Or upload in the Claude apps. If you use the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 4Confirm you have a second model. Make sure at least one other AI model provider is configured with its API key, because the review cannot run without one.
- 5Ask for a review in plain language. Request a second opinion, for example "double check this auth flow change", on work that is significant enough to justify it.
- 6Read the report and decide. Review the agreement and disagreement report, then approve or reject each finding yourself.
What the Cross-Modal Review skill makes the agent do
When the skill is installed and invoked, the agent passes its work product to a different AI model for review. The work product can be a code diff, a brain page, an analysis or a decision. The reviewer grades the work against the original promise, called the Contract, for that work, and reports pass or fail with specific citations.
The agent then presents where the two models agree and where they disagree. The skill is explicit that reviewer suggestions are informational. The agent does not apply them until you approve each one.
- Captures the work to be reviewed (diff, page, analysis or decision).
- Loads the Contract of the originating skill, meaning what was promised.
- Sends the work and Contract to a different model chosen via the project's model-routing conventions.
- Grades the output as PASS or ISSUES FOUND, with evidence.
- Reports agreement and disagreement and waits for your decision.
Install the skill (ZIP download or Settings upload)
Download the skill ZIP from this page and unzip it. For a personal install, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.
If you use the Claude apps instead, you can upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
~/.claude/skills/Trigger the skill with natural requests
You do not need a special command. The skill responds to phrases such as the ones below. Its when_to_use text lists the triggers "second opinion", "cross-modal review", "double check this", "get another perspective" and "challenge this code". A separate adversarial review trigger also exists, "adversarial review".
The skill decides whether a review is worth running. It recommends invoking the review for significant changes, and advises against it for trivial ones.
- "Get a second opinion on this refactor before I commit."
- "Double check this auth flow change."
- "Challenge this code for security problems."
- "Do an adversarial review of the webhook handler."
- "Get another perspective on this migration plan."
When the skill should and should not run
The skill lists situations where a review is appropriate: commits touching 5 or more files or 100 or more lines, security-sensitive changes such as auth flows and webhook transforms, being stuck after 2 or more iterations, before bulk operations or migrations, and when creating or modifying skills.
It says not to run for simple memory writes, single-file typo fixes, routine cron or heartbeat output, or git commit and push of work that was already reviewed.
- Good fit: significant code changes, security-sensitive changes, pre-bulk-operation checks, skill creation or modification.
- Poor fit: typo fixes, formatting changes, routine memory writes, already-reviewed commits.
Use the Codex code-review handoff for diffs
For diff review, the skill can recommend gstack's /codex skill, which wraps the OpenAI Codex CLI. The user invokes /codex review. The skill recognises when that tool is the right choice, for example after a substantial diff lands and before merge, when you want a second opinion that is not another Claude model.
There is also an Adversarial Challenge mode for security-sensitive changes. The reviewer looks for injection vectors, race conditions, auth bypasses, data leaks and privilege escalation paths, and adds an exploitability rating (CRITICAL, HIGH, MEDIUM or LOW) with recommended mitigations.
- Codex Review: independent diff review from a different AI system.
- Adversarial Challenge: security-focused review with exploitability ratings.
- Output lists overlapping findings, findings only Codex made and findings only the agent made, plus an agreement percentage.
/codex reviewUnderstand refusal routing and the user's final say
If the primary review model refuses, the skill switches silently to the next model in the chain defined in conventions/cross-modal.yaml, without showing the refusal or announcing the switch. It escalates to you only if every model in the chain refuses.
The skill also states that cross-model agreement is a signal, not a final verdict. Nothing from the reviewer is applied without your explicit approval, even when the reviewer is right.
- Refusal from one model: next model is tried silently.
- All models refuse: the agent escalates to you.
- Reviewer findings are informational until you approve each one.
What people use it for
- Getting a second opinion on a large refactor before merging to main.
- Checking an authentication flow or webhook handler for security problems with adversarial review.
- Validating a planned batch enrichment or data migration before running it.
- Reviewing a newly written or rewritten skill that changes how the agent behaves.
- Checking a brain page against the Contract of the skill that produced it.
Frequently asked questions
How do I install the cross-modal-review skill?
Download the skill ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
Does the cross-modal-review skill require an API key?
Yes. The skill needs access to at least one other AI model provider, using API keys for the configured reviewer models. Without a second model provider the review cannot run, and it must not be reported as passed.
Is the cross-modal-review skill free?
The source material lists the licence as MIT. It does not state any price, so check the repository for current terms and any costs your model providers charge for API use.
Does the skill run scripts or use the network?
The listing says it runs no scripts and does use the network, because it sends work to other model providers. The optional Codex handoff also requires the gstack /codex skill and the OpenAI Codex CLI.
What if the review model refuses to answer?
The skill switches silently to the next model in its configured chain and does not show the refusal. It only escalates to you if every model in the chain refuses.
Is cross-model review more accurate than a single model?
The skill presents cross-model agreement as a signal, not proof. It does not promise a correct verdict, so you still need to judge the findings yourself.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you want an independent second opinion on significant code changes, security-sensitive work or bulk operations, and you can give it access to a second model provider. Skip it for typo fixes and routine memory writes, and if you have no second model available, since the review cannot run in that case.
Download the skill