---
name: cross-agent-review-loop
description: Coordinate a coding task between separate AI agents using explicit plan review and implementation review handoffs. Use when a task has meaningful design or code risk and the user can ask another agent to review it.
license: CC-BY-4.0
---

# Cross-Agent Review Loop

Use this skill to move one coding task through a plan, an independent review, implementation, and a second review. It works as a set of instructions in a coding agent. It does not connect agents or call another model by itself.

## Use it when

- The task changes several files, affects an important workflow, or has a meaningful chance of causing a regression.
- A second agent can inspect the same project, or the user can safely pass a short handoff between tools.

For a tiny, obvious change, keep the workflow lightweight or skip the second review.

## Rules

- Read the repository's instructions and current state before proposing changes.
- Keep the planning and review stages read-only. A reviewer reports findings and does not edit the project.
- Never claim a second agent reviewed files it could not access. Distinguish a full project review from review of a copied plan or diff.
- Ask the user to transfer the handoff between tools unless a real, approved agent-to-agent connection is already available. Do not invent or silently configure an API, plugin, MCP server, or automation.
- Before sharing a handoff outside the current tool, remove credentials, personal data, private customer information, and unrelated source. Ask the user before sharing code they have not authorized to leave the current environment.
- A review is evidence, not approval. The user remains in charge of scope, permissions, and final decisions.
- Do not delete, publish, deploy, send messages, spend money, or change access as a side effect of this skill without the required authorization.

## Workflow

1. **Restate the task.** Record the requested result, acceptance conditions, constraints, and which agent can access the project. Ask only for missing details that materially change the plan.
2. **Inspect and plan.** Read relevant project guidance, inspect the current files and changes, then produce a concise plan. Include likely files, implementation steps, checks, risks, and unresolved questions. Do not edit during planning.
3. **Prepare a plan handoff.** Use `references/handoff-template.md` or provide the same fields in chat. Give the second agent the task, relevant constraints, and the plan. Ask it to review only, cite evidence by file and line where possible, and avoid edits.
4. **Resolve plan findings.** Sort findings into blockers, important risks, and suggestions. Correct factual mistakes. Ask the user to decide when a finding changes scope or there is a real design tradeoff. Repeat the review after meaningful plan changes.
5. **Confirm implementation authority.** Proceed only when the user's request authorizes the edits and the plan is settled. If that authority is unclear, show the plan and ask before changing files.
6. **Implement and verify.** Make the smallest complete change, preserve unrelated work, and run the checks required by the project or requested by the user. Report checks that could not run.
7. **Review the result.** Give the reviewer the accepted plan, concise diff or changed file list, and verification results. Ask for a read-only review of correctness, regressions, and unmet acceptance conditions. State whether it could read the actual project or only the handoff.
8. **Close the loop.** Fix supported findings within the user's authorized scope, rerun relevant checks, and give the user the final result plus any unresolved risk. If reviewers disagree, explain the evidence and ask the user to choose.

## If a second agent is unavailable

Offer a paste-ready handoff the user can take to another tool. If the user chooses to continue with one agent, label its review as a self-review, not an independent review. Do not pretend the second role ran.

## Expected handoff

A useful handoff can be read on its own. It contains the goal, acceptance conditions, relevant repository instructions, plan or diff summary, exact questions for the reviewer, and known checks. Do not include secrets or unrelated conversation history.

## Attribution

Created by Nishant Gupta / [Code With Nishant](https://codewithnishant.dev/). This pack adapts the Claudex plan-review-build-review workflow into manual, portable handoffs for coding agents.

Licensed under CC BY 4.0. You may share and adapt it with attribution. See `../LICENSE.txt`.

Claudex is an independent workflow guide. It is not an Anthropic or OpenAI integration, and it does not imply endorsement by either company.