meta-research

What This Skill Does

The Meta-Research agent is an autonomous workflow orchestrator designed to transform the chaotic process of academic and technical research into a rigorous, audit-ready pipeline. It functions as a research copilot that enforces a reproducibility-first methodology. Unlike standard brainstorming tools, Meta-Research maintains a state machine across five key phases: brainstorming, literature review, experimental protocol design, data analysis, and documentation. It manages complex projects through a structured directory system that forces documentation of decisions, alternatives, and rationale in real-time. By utilizing a 'logbox' tracking mechanism, the agent ensures that every shift in research direction is captured, preventing the loss of context that often occurs during pivots.

Installation

To integrate this agent into your OpenClaw environment, execute the following command in your terminal or command interface:

clawhub install openclaw/skills/skills/amberljc/meta-research

Use Cases

• AI Research & Benchmarking: Systematically design experiments to test scaling laws or new model architectures while keeping environments pinned.
• Scientific Literature Reviews: Synthesize vast amounts of academic papers into evidence maps with clear citations and critical evaluations.
• Complex Data Analysis: Execute reproducible analysis pipelines where raw data is immutable and every processing step is logged.
• Technical Documentation: Transition from raw research notes to peer-ready manuscript drafts using the agent's built-in writing phase.

Example Prompts

1. "I want to research the effects of different tokenization strategies on model performance; start a new project in the meta-research pipeline."
2. "Review the current literature on transformer attention mechanisms and help me brainstorm three novel ways to optimize memory usage."
3. "My latest analysis of the 001-scaling-laws experiment yielded unexpected results; help me backtrack to the protocol design phase to check for confounding variables."

Tips & Limitations

• Start Simple: The agent supports lazy initialization; you don't need to create complex directory structures immediately. It will prompt you to restructure when you reach your first pivot.
• Embrace Backtracking: The system is designed to handle failure gracefully. If an experiment fails, keep the files in place, mark them as 'archived' in the LOGBOX.md, and start a new folder in explorations/.
• Falsification Focus: Encourage the agent to try and disprove your hypothesis. A rigorous design focuses on identifying why an idea might be wrong, rather than just seeking confirmation. Avoid treating the agent as a 'yes-man' by specifically asking it to seek out contradictory evidence.