Research Skills

Research Skills is a focused collection of agent skills for running an end-to-end research loop: grounding a research question, selecting and executing decisive experiments, writing only from accepted evidence, and reviewing the resulting paper before choosing the next step.

The loop is coordinated by auto-research-orchestrator:

1. Discover grounds the problem in user intent and closest prior work.
2. Experiment selects work by paper-decision value, runs real systems, and independently reviews the result.
3. Write propagates accepted evidence into the paper without changing the scientific contract.
4. Review checks citations, structure, terminology, claim-evidence alignment, and the strongest reject arguments before routing the next step.

Skills

Orchestrate

• auto-research-orchestrator: coordinates the full EXPERIMENT → WRITE → REVIEW lifecycle and preserves RQs, evidence, and research state.

Discover

• research-literature-novelty: builds a claim-oriented literature and novelty map.
• iter-refine-ideas: discusses scientific framing and architecture through read-only research rounds.

Experiment

• research-experiment-design: selects, plans, reviews, executes, and interprets one highest-paper-value real experiment.
• paper-figures: creates and audits design diagrams and result figures tied to specific claims.

Write

• iter-refine-writing: runs structured review-fix rounds over a systems paper.
• rewrite-abstract-intro: restructures and rewrites the abstract and introduction.
• rewrite-paper-section: rewrites a selected body section while preserving technical meaning and evidence.
• paper-writing-style: checks sentence-level systems-paper prose.
• tighten-prose-systems-latex: removes redundancy while preserving mechanisms, qualifiers, and citations.

Review

• check-paper-citations: verifies that citations are real and support the associated statements.
• check-paper-structure-flow: checks section organization, paragraph roles, and sentence flow.
• check-terminology-infoflow: checks terminology, concept consistency, and information flow.
• iter-review-critique: performs iterative adversarial review of a complete paper.

Core Principles

• Paper-decision value over activity: more runs, reports, or checker passes do not automatically constitute research progress.
• Real evidence over proxies: use real systems, official benchmarks, complete runs, raw artifacts, and reproducible analysis paths.
• One-way scientific authority: user intent defines the objective; literature, experiments, and reviews provide evidence; writing expresses accepted decisions.
• Independent result review: smoke tests and partial runs cannot become paper evidence without completing and reviewing the planned experiment.
• Persistent, recoverable state: preserve RQs, experiment plans, raw results, paper decisions, and the reason for the next action.

Installation

Clone the repository and copy or symlink the desired skill directories into your agent's skill directory. For Codex:

git clone https://github.com/eunomia-bpf/research-skills.git
mkdir -p ~/.codex/skills
cp -R research-skills/skills/* ~/.codex/skills/

Each skill is self-contained under skills/<skill-name>/ and starts with a SKILL.md file. Some skills reference skill-creator or evolve-agent-skills as optional external integrations; those general-purpose meta-skills are intentionally not bundled here.

Repository Layout

research-skills/
├── README.md
├── docs/
│   ├── research-skills-loop.drawio
│   └── research-skills-loop.png
└── skills/
    └── <skill-name>/
        ├── SKILL.md
        └── references/ or scripts/

The editable loop diagram is available at docs/research-skills-loop.drawio.