Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    AI-Loop — Independent Verification for Claude Code

    1

    A gated multi-agent workflow for Claude Code with adversarial review, independent verification, and evidence before human acceptance.

    $8

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    AI-Loop — Independent Verification for Claude Code

    AI-Loop — Independent Verification for Claude Code

    Example session with this skill installed

    Use AI-Loop to let a tournament organizer replace a team with a team from another club. Test the feature in the web app using Chrome.

    • Read your context and instructions
    • Compiled the ai-loop

    External-team replacement — implementation and verification report

    Translated and anonymized summary of an actual Claude Code session. This is an edited case study, not a verbatim transcript or a new execution.

    The feature was implemented across the backend, web, and mobile repositories. An organizer can select a replacement from other clubs; the incoming team is invited, and its matches remain blocked until it accepts.

    Review caught issues before handoff

    The final cross-repository review found that failed replacements did not display an error in the web app and that an error message was missing. Both issues were corrected, and the follow-up review returned clean.

    Verification evidence

    CheckRecorded result
    Eight implementation tasksPassed their reviews
    Backend acceptance test run67 PASS, 0 FAIL, 0 UNVERIFIED
    Baseline comparisonThe pre-feature binary failed the expected new replacement and consent checks; perimeter checks remained unchanged
    Backend regression suiteMatched the baseline package by package
    Authorization checksPassed with zero authorization orphans
    Affected web test directoriesPassed after the final corrections

    End-to-end flow exercised in Chrome

    1. Open the replacement dialog and select a team from another club.
    2. Confirm that the UI warns that the incoming team requires acceptance.
    3. Complete the replacement and confirm that the incoming team appears as invited in the registration inbox.
    4. Attempt to start its match before acceptance: the app blocks the action and names the team.
    5. After acceptance, start the match successfully.

    The local browser trace contains navigation, clicks, and screenshots for the flow. Acceptance in the E2E scenario used a synthetic session; this does not establish a complete two-user acceptance UI walkthrough.

    Declared limits and handoff

    • The complete web suite was not rerun after the last corrections; affected directories were rerun.
    • The new blocking message was covered by a mobile widget test but was not inspected on a physical phone.
    • The match-start E2E used a prepared tournament fixture and did not cover the creation wizard.
    • Branches were ready for owner review. Integration with newer changes, deployment checks, and final acceptance remained separate steps.

    The handoff delivered the implementation, verification results, browser-tested behavior, and remaining limitations for the human owner to review.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Your AI agent says “done.” What proves it?

    When the same agent plans, implements, and checks a change, its blind spots can survive the review. AI-Loop gives Claude Code a structured multi-agent protocol: separate roles, explicit quality gates, and evidence for every acceptance criterion.

    The implementer does not verify its own work. Passing tests does not automatically mean acceptance. The human owner decides when the work is done.

    What it does

    • Separates implementation from verification. Distinct agents handle adversarial review, implementation, and independent verification.
    • Challenges the specification. An adversarial gate tests assumptions and technical claims before implementation proceeds.
    • Requires meaningful test evidence. The implementation protocol calls for failing assertions, baseline checks, and mutation checks where applicable.
    • Verifies independently. A separate verifier uses its own probes on a different model. The coder’s report is an input, never proof.
    • Makes uncertainty explicit. PASS requires evidence for every acceptance criterion. FAIL, BLOCKED, and UNKNOWN have distinct routes; UNKNOWN never counts as PASS.
    • Structures handoffs. Written handoffs and fresh sessions at phase boundaries help manage context and preserve progress.

    How it works

    1. Versioned specification: Define scope, acceptance criteria, and measured technical claims.
    2. Adversarial gate: Challenge the plan using independent checks and isolated worktree probes.
    3. Implementation: Execute the approved specification and report changes, checks, and unresolved limitations.
    4. Independent verification: Re-measure acceptance criteria and review the complete diff.
    5. E2E checks: Verify the resulting behavior through the relevant end-to-end checks.
    6. Human acceptance: Present the evidence and remaining risks for the owner’s decision.

    Best suited to

    • Root-cause fixes that need more than a plausible patch.
    • Features involving schemas, permissions, data, or explicit behavioral contracts.
    • Substantial refactoring where existing behavior must be preserved.
    • Changes that benefit from adversarial review and clear approval boundaries.

    What you get

    The complete SKILL.md, a quick-start installation guide, and the original MIT license. Install it as a personal or project skill in Claude Code.

    Requirements and limitations

    Requires Claude Code with skills and subagents, access to a different model for independent verification, and a Git repository with the relevant build, test, and E2E tools. Human participation is required for approvals and session handoffs. Model usage costs are separate.

    AI-Loop provides agent instructions, not deterministic enforcement. Compliance and bug-free results are not guaranteed. The workflow adds time and token overhead and is not suitable for trivial renames or formatting changes.

    No model access, executable engine, automated test fixtures, or personalized consulting is included. Compatibility with other agents has not been validated.

    Open source, packaged for convenience

    AI-Loop is available free under MIT on GitHub. Purchase this ready-to-install package through Agensi to support the author.

    Requires Claude Code with skills and subagents, plus access to a different model for independent verification. Adds time and token overhead; costs vary by task and configuration. Not suited to trivial renames or formatting changes.

    The protocol uses written handoffs and fresh sessions at phase boundaries, which may require manual continuation. It provides agent instructions, not deterministic enforcement, and does not guarantee bug-free results. Build, test, browser, and E2E tools must be available in the user's environment. Other agent platforms have not been validated.

    How to use

    1. 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. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    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 1 month ago

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Browser
    Network Access
    Environment Variables

    Reads project instructions and source files, writes versioned specifications and implementation changes, and runs build/test commands and isolated Git worktree probes. File access is limited to the repository and paths authorized for the task. Browser and network access are conditional on the project's E2E and research needs. Environment variables are used only for authorized toolchain and test configuration; the workflow does not require collecting or exposing secrets. Production data, destructive actions, infrastructure changes, and deployment require explicit owner authorization under the protocol. Final acceptance remains with the human owner.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    AI-Loop introduces a 'separated duties' protocol where the agent that writes the code is not the one that verifies it. It solves the problem of 'hallucinated success' by requiring independent adversarial review and evidence-based quality gates before any code is presented to you for final approval.

    This skill is specifically designed for Claude Code and utilizes its subagent capabilities. For the independent verification phase, it is recommended to have access to a secondary model (such as a different version of Claude or a GPT-4 variant) to ensure a truly unbiased review process.

    You will receive the complete SKILL.md file, a quick-start installation guide, and the MIT license documentation. This package provides the structured prompts and workflow instructions needed to instantly add these multi-agent capabilities to your Claude Code environment.

    AI-Loop is best reserved for complex tasks like root-cause bug fixes, schema changes, or major refactors where errors are costly. Because it involves multiple steps and independent verification, it is not recommended for trivial tasks like typo fixes or CSS formatting where the overhead would outweigh the benefits.

    Installation is straightforward: you add the provided SKILL.md to your project's skill directory or your global Claude Code configuration. The included guide walks you through the folder structure and how to trigger the 'Adversarial Gate' to begin a new verified session.

    While the core logic is provided as a ready-to-use skill, it is licensed under MIT, allowing you to tweak the quality gates or reporting templates to match your specific team's coding standards or testing frameworks.

    New in Agents & Orchestration