Autoresearch Paper: Codex + Claude Code ML Research Pipeline

    1

    For AI/ML researchers running Codex + Claude Code: turns a research brief into a gated pipeline that freezes the evaluator first and blocks paper writing until the experiment actually passes.

    $9.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Autoresearch Paper: Codex + Claude Code ML Research Pipeline

    Example session with this skill installed

    Here's my research brief: energy-efficient coverage path planning for UAV swarms in wind-affected environments, with a learned wind-compensation policy. Target venue: CVPR 2027. My reference PDFs are in ./uav-refs/. Set up the research plan before running anything.

    • Read your context and instructions
    • Compiled the autoresearch paper

    Illustrative walkthrough, not a recorded run. The package ships no real end-to-end run for the main skill, so this describes what the skill does with this brief and shows no invented results.

    1. Brief check: the skill parses the paragraph into its closed brief. "CVPR" matches the conference tier, so it proposes the conference tier for you to confirm. It then lists the required fields it could not infer safely (for this brief: candidate ideas, code location, initial direction, the strongest comparable baseline, the evaluator/metric context, resource bounds, permissions and stop conditions) and asks for them. No plan, model call, schedule or file is created until the brief is complete.
    2. Plan: once the brief is complete, it prepares a conference-tier plan in which the evaluator is frozen first (T0), followed by literature and gap analysis, method, implementation and experiments, then evaluate-candidate, research-decision, pivot-or-retry and figure-build steps. Writing comes last.
    3. Authorization and review: you sign one authorization for the plan, and an independent Codex review checks the first stage. The Claude Code worker starts only after both.
    4. Gates while it runs: writing stays blocked until the frozen evaluator records a PASS (or you sign a waiver) and the final-evidence review is approved. Paper figures also need a provenance manifest and your review. You can check status, pause, resume or stop at any time, and stop ends with a cleanup report.

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

    About this skill

    Make your agents prove the experiment works before they write the paper.

    Long agent-driven research runs tend to fail the same way. The agent explores for hours, forgets what it already tried, retries the same weak idea, and then writes a polished paper about a result that never cleared the bar. Autoresearch Paper is a Codex + Claude Code skill that turns this failure into one you can see and recover from.

    What it does

    • Closed brief first. It turns your paragraph-level brief into a closed research brief covering objective, venue tier, baseline, evaluator/metric, resource bounds, permissions and stop conditions. If a required field is missing, it names the field and stops. No plan, model call or schedule is created until the brief is complete.
    • Evaluator frozen up front. The metric and threshold are fixed before any method work starts. Verdicts are hash-bound to the evidence they came from, so the goalposts can't move.
    • Writing gated on evidence. Conference and Q1-journal papers start only after a validated evaluator PASS, or a waiver you sign, plus a final evidence review. The arXiv tier is open.
    • Structural pivots, not endless retries. Failed directions are tracked. Repeated failures in different directions trigger a pivot review instead of small retries.
    • Figure gate. A paper figure needs a provenance manifest and your review before it can be used.
    • One authorization, no silent autonomy. You sign one explicit authorization. An independent Codex review checks the first stage before the Claude Code worker starts. Silence never counts as approval.
    • You stay in control. It never auto-aborts or deletes on its own. You get status, pause, resume, stop, patrol and cleanup of every resource it created, plus a read-only local dashboard (loopback only).

    Who it's for

    AI/ML researchers who already run long Codex or Claude Code sessions, and who can define an evaluator (a metric plus a threshold) for their idea. It is not for one-off drafts, blog posts, slides, or camera-ready formatting.

    Free on GitHub vs. this listing (honest version)

    The core skill is MIT open source: https://github.com/WdBlink/autoresearch-paper. If you are happy to install from GitHub, you can use it for free.

    This $9.99 listing gives you:

    • A curated stable build (v0.20.1). A single skill with the experimental MVP-0 preview branch removed, so your agent loads one clean SKILL.md.
    • One-prompt install through Agensi. Your agent installs it for you, and the package has passed Agensi's security scan.
    • Support for continued development. Purchases fund the more advanced paid version that is in the works.

    Requirements and honest status

    • Host: Codex CLI/App, signed in. Worker: a separate Claude Code session; tested with MiniMax M3 as the worker model. Other agent pairings are not supported yet.
    • OS: macOS recommended, because the independent L0 watchdog uses launchd. Linux and Windows are untested.
    • Status: experimental (v0.20.1). Bounded field acceptance passed. It does not claim 24h/7×24 stability, production readiness, or fully autonomous paper completion.

    What it installs on your machine

    Per-plan launchd jobs (labels com.autoresearch-paper.<hash>) for the watchdog, plus files inside the plan directory you choose. stop and cleanup remove only the resources the plan owns. The legacy MAVIS rescue mode is opt-in only.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Research & Analysis.

    Version 0.20.1 (experimental). MIT licensed. Author: Pengyu Yin. Known limitations: Codex + Claude Code only; macOS-first; no 24h/7×24 stability claim; conference/journal tiers require an evaluator you can define. Changelog and issues: https://github.com/WdBlink/autoresearch-paper

    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 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 28 days ago

    Software
    Node.jsPython 3Codex CLIClaude CodeLaTeX (TeX Live)

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables
    Listed28 days ago

    What's inside

    Frequently Asked Questions

    The core skill is the same MIT-licensed code. This listing is a curated stable build (v0.20.1, single skill, experimental preview removed) that installs in one prompt through Agensi and passed Agensi's security scan. Buying it supports development of the upcoming advanced version.

    No claim of fully autonomous paper completion. It runs the research loop and only allows writing once a frozen evaluator records a PASS or you sign a waiver. You review and sign the key decisions.

    Codex as host plus a separate Claude Code session as worker (tested with MiniMax M3). Other agents are not supported yet.

    Untested. The independent watchdog uses macOS launchd, so macOS is recommended.

    Per-plan launchd jobs for the watchdog, plus files in the plan directory you choose. `stop`/`cleanup` remove only what the plan created. It never deletes anything without your confirmation.

    That depends on your Codex plan and your Claude Code/MiniMax usage. The skill freezes a budget per plan, but model costs are billed by your own providers.

    New in Research & Analysis