Can I use my own evaluation scripts?

Absolutely. You can provide any custom script that prints a metric name and value to stdout, allowing for highly specific optimization workflows.

What happens if the agent gets stuck in a crash loop?

The skill includes a safety protocol that pauses the loop and alerts the user after 5 consecutive crashes to prevent wasting compute cycles on broken logic.

Does this work for non-code files?

Yes, it can optimize anything with a measurable metric, including marketing headlines, system prompts, and documentation using built-in or custom LLM judges.

How does the agent decide which changes to keep?

The agent runs a user-defined evaluation command after every edit. If the resulting metric improves based on your target (e.g., lower latency or higher CTR), it commits the change; otherwise, it performs a git reset.

Is a git repository required to use this skill?

Yes, the agent relies on git to manage the versioning of successful iterations and the 'undo' functionality for failed experiments.

Autoresearch Agent

Name: Autoresearch Agent
Author: alirezarezvani

byalirezarezvani

•

9,958

•

생산성 및 워크플로

Automates repetitive experiment loops to optimize code performance, file sizes, or content quality through continuous measurement and git-backed iterations.

Inspired by Karpathy’s autoresearch, this skill empowers Claude to act as an autonomous optimization agent that works while you sleep. It systematically edits a target file, executes a measurable evaluation command, and uses git to version-control improvements while discarding failures. Whether you are minimizing bundle sizes, speeding up API response times, or refining marketing copy, the Autoresearch Agent runs indefinitely—exploring dozens of variations to deliver a highly optimized result based on hard data rather than guesswork.

주요 기능

01Supports local engineering metrics like speed, size, and test pass rates

02Native Git integration for automatic commits of improvements and resets of failures

03Autonomous experiment loops with configurable intervals (10m to monthly)

04Real-time dashboard to track runs, delta improvements, and experiment status

05Built-in LLM judges for qualitative metrics like headline CTR and prompt quality

069,958 GitHub stars

사용 사례

01Iteratively improving system prompts or marketing copy using LLM-based engagement scores

02Minimizing JavaScript bundle or Docker image sizes for production deployment

03Optimizing backend API performance and reducing latency through systematic refactoring

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add alirezarezvani/claude-skills autoresearch-agent

For use in Claude.ai and ChatGPT

주요 기능

01Supports local engineering metrics like speed, size, and test pass rates

02Native Git integration for automatic commits of improvements and resets of failures

03Autonomous experiment loops with configurable intervals (10m to monthly)

04Real-time dashboard to track runs, delta improvements, and experiment status

05Built-in LLM judges for qualitative metrics like headline CTR and prompt quality

069,958 GitHub stars

사용 사례

01Iteratively improving system prompts or marketing copy using LLM-based engagement scores

02Minimizing JavaScript bundle or Docker image sizes for production deployment

03Optimizing backend API performance and reducing latency through systematic refactoring

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add alirezarezvani/claude-skills autoresearch-agent

For use in Claude.ai and ChatGPT