This Claude Code skill, uditgoenka/autoresearch, automates goal-directed iteration for any task with a measurable outcome, generalizing Karpathy's "constraint + metric + iteration = compounding gains" principle beyond ML. It empowers users to define goals, and Claude autonomously refines code, content, or processes through a loop of modification, verification, and selective retention, driving significant productivity gains.

How It Works

The system operates via an autonomous loop: Review state -> Ideate change -> Modify -> Git commit -> Mechanically Verify -> Decide (Keep/Revert) -> Log -> Repeat. This process emphasizes constraint (single atomic change, defined scope), a clear success criterion (mechanical metric), and fast, automated verification. Git serves as memory, tracking successes and automatically reverting failures, enabling stable, compounding improvements.

Quick Start & Requirements

Install by cloning the repo and copying the autoresearch skill to your Claude Code skills directory (e.g., .claude/skills/autoresearch/). Primary commands include /autoresearch (unlimited loop) and /loop N /autoresearch (N iterations, requires Claude Code v1.0.32+). The /autoresearch:plan wizard assists in defining Goal, Scope, Metric, and Verify commands. A mechanical, quantifiable metric and a fast verification command (<10 seconds) are essential.

Highlighted Details

• /autoresearch:plan: Interactive wizard for defining and validating Goal, Scope, Metric, and Verify commands.
• /autoresearch:security: Autonomous, read-only security audit using STRIDE, OWASP, and red-team personas. Generates detailed reports and supports --diff, --fix, --fail-on flags.
• /loop N: Enables fixed-iteration runs for controlled sprints and CI/CD integration.
• Domain Generalization: Extends beyond ML to software engineering, content, documentation, and any quantifiable task.

Maintenance & Community

Authored by Udit Goenka. No specific maintainer details or community channels are listed in the README.

Licensing & Compatibility

Released under the MIT License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Success hinges on defining a precise, mechanical metric and a fast, deterministic verification command; subjective criteria or slow checks hinder progress. Requires the Claude Code environment. The /autoresearch:plan wizard helps mitigate setup complexity.