- Home
- Skills
- Agents & Orchestration
- Cross Family Gatekeeper
More screenshots
Works with the AI Agents you already use
Cross Family Gatekeeper
Enforce adversarial, cross-model-family review cycles to verify code changes and technical claims.
$7
Cross Family Gatekeeper
Example session with this skill installed
Use Cross-Family Gating Orchestrator to coordinate a focused fix to our record-deduplication function.
Each record has a string identifier. Keep the first record for each exact identifier, preserve first-appearance order, ignore empty or whitespace-only identifiers, and leave the input unchanged.
Assign the builder, independent reviewer and verification responsibilities using our project policy and available authorized agents. The approval reviewer must use a known different model lineage from the builder. Reuse our existing project handoff and review-record locations.
Freeze the candidate and requirements before review. Have the reviewer inspect the actual proposal independently, verify substantive findings with focused checks, and re-review any correction against the updated candidate.
Return the role assignments, proposed function, review findings, verification results and next action. Keep review approval, executed checks and action authorization separate. If independent review or evidence is unavailable, report the incomplete gate.
Do not add unrelated changes, merge or deploy..
- Read your context and instructions
- Compiled the cross family gatekeeper
Focused Task and Gating Report
The revised deduplication proposal preserves the first qualifying record for each identifier and maintains input order. An independent challenge exposed a last-record-wins defect in the initial proposal; the correction stayed within the requested function.
Illustrative completed scenario, not an actual execution log or review receipt. Family A/B and candidates R1/R2 are explanatory labels. A live run records observed model identities, exact candidate hashes and original evidence.
Task assignments
| Responsibility | Assignment and boundary |
|---|---|
| Coordinator | Maintains scope, requirements and the gate record. Follows the project’s approval and authority rules. |
| Builder — Family A | Prepares the proposed function and makes the bounded correction. Does not approve its own work. |
| Independent reviewer — Family B | Inspects the frozen candidate read-only against the same requirements, without other reviewers’ verdicts. |
| Verification responsibility | Checks the revised behavior and records actual results against the tested candidate. |
| Action-owner | Retains permission to apply, merge or deploy. Review approval grants none of those permissions. |
Independent finding and correction
Candidate R1 stored records in a dictionary keyed by identifier. A later duplicate replaced the earlier record, violating the keep-first requirement.
The reviewer demonstrated the issue with two records sharing identifier A but containing different values. Candidate R2 replaces that behavior with a seen-identifier set and appends only the first qualifying record.
Proposed function
def dedupe_records(records):
seen = set()
result = []
for record in records:
record_id = record["id"]
if not record_id.strip() or record_id in seen:
continue
seen.add(record_id)
result.append(record)
return result
Verification results in this worked scenario
| Check | Result |
|---|---|
Two records with identifier A | First record retained. |
Identifiers A, B, A | First A, then B, preserving order. |
| Empty or whitespace-only identifier | Record skipped. |
| Empty input | Empty result returned. |
| Input list and original records | Unchanged. |
A real run attaches the executed command, its output and the exact tested candidate. This illustrative table is not a substitute for that evidence.
Gate and action state
- Review: The revised review covers R2 and the unchanged requirements. R1’s receipt remains historical.
- Verification: Focused checks pass in this illustrative scenario; a live run retains the execution evidence.
- Authorization: No merge or deployment is authorized by this report.
- Scope: No unrelated refactor, new framework or additional feature was introduced.
Next action
Present R2 and its supporting gate record to the authorized owner. If the candidate or relevant requirements change, refresh the affected checks and independent review before reporting a new pass.
Connects securely to your tools. The creator never sees your data.
About this skill
Consequential code changes often bypass rigorous review when handled by solo agents. This skill enforces a structured cross-family adversarial review cycle to ensure technical claims and diffs are verified by independent model lineages before proceeding.
What it does
- Coordinate cycles by managing the propose, independently review, verify, and decide workflow for code milestones.
- Freeze candidates by pinning content hashes and snapshots to ensure reviewers evaluate the exact code intended for change.
- Verify independence by identifying model families from runtime metadata to prevent "self-review" by the same model lineage.
- Audit evidence by classifying findings as confirmed, unresolved, or disputed based on reproducible counterexamples and source spans.
- Manage handoffs by tracking gate status, execution signatures, and remaining conditions across different agent sessions.
How it works
- Establish task by reading project instructions to identify authorized decision-makers and the required gating policy.
- Snapshot candidate to capture the diff, baseline, and test output, ensuring the reviewer sees the raw artifacts rather than summaries.
- Dispatch review to a different model family with specific instructions to find requirement violations or material risks.
- Reconcile findings by verifying reviewer claims against the code and reporting the final gate status to the project lead.
Frameworks & tools
This skill is designed for multi-agent environments using diverse LLM providers (e.g., Anthropic, OpenAI, Google). It works alongside Git for diff tracking and standard test runners for evidence collection.
Why this beats prompting it yourself
Standard prompts often let agents "rubber stamp" their own work or accept vague summaries. This skill enforces strict identity checks between model families and requires raw evidence for every claim, preventing silent failures in the review chain.
Use cases
- Gating pull requests by requiring a PASS from a different model family before merging.
- Validating architecture plans to find logic flaws before implementation starts.
- Recovering review context after a session handoff by verifying candidate hashes.
- Auditing high-risk changes where a single agent's bias might overlook edge cases.
Known limitations
Does not provide a sandbox or model service. Cannot verify independence if model metadata is hidden or falsified by the provider. Requires an existing project policy to define decision-makers.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires access to model metadata to verify family independence. Does not automate model deployment or cost management. Scope is limited to technical review, not legal or financial approval.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 14 days ago