agent-research-skills for Claude Code

The github-research skill is designed to systematically discover, evaluate, and analyze GitHub repositories related to a specific research topic. It integrates with deep-research output (such as paper databases and phase reports) to produce an actionable integration blueprint for reusing open-source code.

Key Workflow:

1. Input: Requires a deep-research output directory containing paper_db.jsonl, phase reports, and code_repos.md.
2. Output: Generates a structured directory at ./github-research-output/{slug}/ containing a master repository database, analysis reports, and integration plans.
3. Pipeline: Operates through a 6-phase process: Intake, Discovery, Filtering, Deep Dive, Analysis, and Blueprinting.

The experiment-code skill is designed to generate, improve, debug, and plot Machine Learning experiment code for research papers. It follows an iterative workflow that emphasizes planning, self-contained code, and proper logging.

Available Actions

Input Arguments

• $0 (Task): One of generate, improve, debug, or plot.
• $1 (Context): A research plan, idea description, or an error message.

The symbolic-equation skill uses LLM-guided evolutionary search (LLM-SR) to discover interpretable scientific equations from datasets. It employs a multi-island algorithm with softmax-based cluster sampling, island resets to maintain diversity, and LLM-proposed mutations to evolve equations.

Input Requirements:

• $0: A dataset description, variable names, and the physical context of the data.

argument-hint: [data-and-variables]

The algorithm-design skill is used to formalize methods into algorithmic pseudocode and system architecture diagrams. It is specifically designed for formalizing methods for research papers, generating LaTeX environments, and creating Mermaid UML diagrams (class and sequence diagrams).

Input:

• $0 — A method description or an existing implementation that needs to be formalized.

algorithm-design [method-description]

The survey-generation skill is designed to produce complete academic survey papers. It utilizes a multi-stage pipeline inspired by AutoSurvey, featuring multi-LLM parallel outline generation, RAG-based subsection writing, citation validation, and local coherence enhancement.

Input:

• $0: The survey topic or research area.

Core Rules:

• No Hallucinations: Only cite papers from the provided collected list; never invent citations.
• Mandatory Validation: Citation validation must be performed before final output.
• Structure: Each subsection must meet a minimum word count and avoid duplicate content.
• Coherence: Local coherence enhancement must be applied while preserving all citations.

The figure-generation skill generates publication-quality scientific figures (bar charts, line plots, heatmaps, etc.) using a three-phase pipeline involving query expansion, code generation with an execution loop, and VLM-based visual refinement.

Arguments:

• $0 (Required): A description of the desired figure.
• $1 (Optional): A path to a data file (CSV, JSON, NPY, PKL) or a results directory.

The self-review skill automatically reviews academic papers (PDF or .tex files) using a structured NeurIPS-style review form. It employs three distinct reviewer personas (Harsh but fair, Harsh and critical, and Open-minded), performs ensemble scoring with reflection refinement, and generates an actionable report with strengths, weaknesses, and specific improvement suggestions.

Usage Pattern: Provide the path to your paper as an argument.

Input:

• $ARGUMENTS: Path to a PDF file or a .tex file.

self-review [pdf-or-tex-file]

The literature-search skill allows you to search multiple academic databases (Semantic Scholar, arXiv, OpenAlex, and CrossRef) to find relevant papers, check related work, or build bibliographies. It returns structured JSONL data containing titles, authors, years, venues, abstracts, citations, and BibTeX.

Input:

• $ARGUMENTS: A natural language search query.

The research-planning skill is used to design research plans and paper architectures. Given a research topic, idea, or an existing paper to reproduce, it generates structured plans including methodology outlines, paper structures, dependency-ordered task lists, and experiment designs.

Input Argument:

• $0 — A research topic, idea description, or a paper to reproduce.

research-planning [topic-or-idea]

The slide-generation skill converts a completed LaTeX paper into presentation slides (using Beamer LaTeX) or a poster. It extracts key figures, tables, and equations to create a narrative flow suitable for oral presentations.

Input:

• $0: A Paper LaTeX file (main.tex) or a paper directory.

AgentLaboratory provides mechanisms for reflecting on code performance and errors to improve future iterations.

General Reflection: Used to derive insights from multiple code variants and their scores: {code_variants_with_scores}.

Error Reflection: When code fails, the agent is prompted to provide a detailed line-by-line analysis of the error and a plan for a fix using EDIT commands, rather than providing entirely new code blocks.