Sw Autoresearch

What This Skill Does

The Sw Autoresearch skill is an advanced agentic tool designed to automate complex investigation workflows. Inspired by Andrej Karpathy's iterative research methodology, this skill transforms the AI agent from a simple question-answer engine into a goal-oriented investigator. It forces the agent to define strict success criteria before it begins, ensuring that every research action is purposeful. The agent executes a recursive loop: generating a hypothesis, executing the research, verifying findings against your predefined goals, and iteratively refining its approach if the results fall short. This process continues automatically until the success criteria are fully met or a safety threshold is hit.

Installation

To add this skill to your OpenClaw environment, execute the following command in your terminal:

clawhub install openclaw/skills/skills/amdf01-debug/sw-autoresearch

Ensure your agent has the necessary permissions to access the internet and read/write scratchpad files to maintain the iteration logs.

Use Cases

• Market Analysis: Gathering data on competitors to meet specific financial or operational metrics.
• Technical Due Diligence: Investigating library documentation or security vulnerabilities where initial answers may be incomplete.
• Academic Inquiry: Synthesizing complex topics that require cross-referencing multiple sources and verifying claims.
• Comparative Reviews: Analyzing multiple software packages based on specific feature requirements provided by the user.

Example Prompts

1. "Autoresearch: Evaluate the performance difference between Rust and Go for real-time data processing. Goal: Identify latency bottlenecks. Criteria: Must compare at least three independent benchmarks, define memory footprint, and list concurrency models."
2. "I need a deep dive into the current state of decentralized identity standards. Research this thoroughly, ensuring at least three distinct sources are cited and cross-verified for any conflicting information."
3. "Iterate until complete: Compare the API capabilities of OpenAI, Anthropic, and Google Gemini for long-context RAG. Success requires a feature matrix, pricing comparison, and maximum token window verification."

Tips & Limitations

• Be Specific: Success criteria are the most important part of this skill. Vague criteria result in vague output.
• Monitor Loops: The skill enforces a 10-iteration limit. If your request is overly broad, the agent may exhaust its iterations before finding an answer.
• Verify Sources: While the agent is programmed to verify, always inspect the "Iteration Log" generated by the skill to understand the agent's logic path and identify potential blind spots.
• Differentiation: The agent must be instructed to vary its approach. If it fails a search using keywords, suggest that it look for whitepapers or community discussions in subsequent attempts.