skill

research-kit

Research Kit organizes AI-assisted research into 6 core stages and 4 domain modules. The workflow verifies sources to keep fabricated references out of the literature record, supports research-questio

vinhnt21/research-kitcommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/vinhnt21/research-kit itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

Research Kit: Lean & Reproducible Research Skills for AI Agents

A lean, reproducible suite of research skills guiding AI agents through the complete scientific paper lifecycle—with empirical rigor, zero bloat, and <0.50% standing context overhead. The kit comprises a 6-skill core pipeline and 4 specialist extensions.

🎉 News & Updates

[2026.10.20] 🎯 [Academic Visuals v2.4.0]: Enhanced slide visualization — new content slides now record an explicit, source-backed visual decision (prioritizing empirical charts and code-generated diagrams over decorative icons, while preserving justified text-only slides). Adds a per-asset license and credit ledger (visual-ledger.json), a safe SVG-to-PNG workflow, and an automated visual-plan linter that flags dense or consecutive text-only slides for review. See the visual asset guide.
[2026.10.04] 🎯 [LaTeX Diagrams v2.3.0]: Improved complex diagram layouts with spacing based on actual block heights, clearer routing and labels, and complete container bounds. Render, inspect and repair before delivery. Includes a runnable two-pass example.
[2026.10.04] 🎯 [Slides & Native Math]: Improved slide rendering with editable equations compiled from LaTeX to native OMML, without special-character substitutes. Supports inline math, matched font sizes and OOXML bounds checks. See math and slide tooling.
[2026.10.03] 🎯 [Dynamic Language Adaptation]: Upgraded all 10 procedural skills to automatically interact and synthesize in the researcher's active language with natural peer-colleague terminology ("như trao đổi cùng đồng nghiệp"), while strictly preserving English instructions for execution consistency and prompt stability.
[2026.10.02] 🎯 [Release v1.0.0]: Formalized Research Kit v1.0.0 comprising 6 sequential scientific lifecycle skills (rk-survey through rk-report) and 4 specialist modules, reinforced by automated SHA-256 cryptographic check-suite verification (scripts/check-suite.py).
[2026.10.01] 🎯 [Documentation & Portal]: Launched dual-language academic web portal (research-kit.vinhnguyenthanh.com) and comprehensive operational guides (documents/guide.md & documents/guide.vi.md).

News & Updates
Overview & High-Level Architecture
- Part 1: Core Pipeline — 6 Skills for the Scientific Paper Lifecycle
- Part 2: Specialist Extensions — 4 Domain-Specific Skills (Author's Research Focus)
Prerequisites & Toolchain Setup
Installation & Quick Start
Documentation
Comparison & Key Advantages
- Direct Comparison: Scientific Agent Skills (v2.65.0) & Science Superpowers vs. Research Kit
- Key Advantages & Architectural Principles
Companion References & Field Templates
References & Prior Art

Overview & High-Level Architecture

Research Kit is an open-source suite of procedural research skills designed to guide AI coding and research agents (Cursor, Claude Code, Codex, Antigravity) through end-to-end scientific paper development. From literature discovery and hypothesis formulation to experimental protocol freeze, fresh data rerun verification, and manuscript drafting, Research Kit enforces empirical rigor at every stage.

The kit is organized into two parts:

Core Pipeline (6 skills): Covers the complete scientific paper lifecycle — from literature survey through dissemination — forming a deterministic, stage-by-stage execution pipeline.
Specialist Extensions (4 skills): Domain-specific modules reflecting the author's research interests (quantum computing, quantum networking, AI/ML, and academic presentations). These are provided as references and can be adapted or replaced to suit your own research domain.

Unlike large catalogs such as Scientific Agent Skills (v2.65.0, which packages 163 uncurated skills consuming 14,246 standing tokens before execution begins), Research Kit consumes less than 0.50% of a standard context window.

Part 1: Core Pipeline — 6 Skills for the Scientific Paper Lifecycle

These 6 skills form the backbone of Research Kit, covering every stage from literature survey to dissemination in a deterministic, sequential pipeline:

Skill	Stage	Primary Responsibility	Mandatory Quality Gate
`rk-survey`	01. Literature Survey	Systematic search, citation validation, gap analysis	Evidence boundary locked & source provenance check
`rk-idea`	02. Idea & Hypotheses	Research question framing, rival explanations, falsification criteria	Pursue / Kill decision matrix before coding
`rk-method`	03. Method & Protocol	Experimental design, controls, power analysis, compute budget	Protocol freeze (prevents post-hoc p-hacking)
`rk-data`	04. Data & Analysis	Exploratory inspection, statistical testing, publication figures	Mandatory fresh rerun check before claiming
`rk-write`	05. Drafting & Revision	Structured drafting, 1 idea/paragraph, reviewer rebuttal	100% claim-to-evidence alignment audit
`rk-report`	06. Dissemination	Progress briefings, executive summaries, defense presentations	Structured reporting gate (grounded claims only)

Part 2: Specialist Extensions — 4 Domain-Specific Skills (Author's Research Focus)

These skills are tailored to the author's research domains and serve as practical references. You can adopt, adapt, or replace them to match your own specialization:

Skill	Domain	Primary Responsibility	Mandatory Quality Gate
`rk-quantum`	Quantum Computing	Local Hamiltonian/circuit simulation & variational algorithms	Local simulation default; QPU spend requires approval
`rk-quantum-network`	Quantum Networking	Entanglement distribution, repeater memory, routing & scheduling	Fidelity checks & protocol verification
`rk-ai`	AI / Machine Learning	Train/val/test splits, leakage audit, baselines & generative eval	Zero data leakage & reproducible seed verification
`rk-academic-visualize`	Scientific Visualization & Slides	Scientific illustrations and source-grounded slide decks with visual decisions and licensed asset provenance	Render and inspect figures; check deck structure, visual intent, and asset credit

🛠️ Prerequisites & Toolchain Setup

Core Markdown SOPs need only a compatible agent. Verification, slide math, and figure rendering use the tools relevant to that task:

Tool	Minimum Version	Purpose	Necessity
Python	3.9+	Test suites, OOXML and visual-plan checks, SVG sanitization, suite hash verification	Required for these checks (uses the standard library; zero external pip bloat)
Pandoc	3.0+	Compiles LaTeX mathematical formulas into native PowerPoint OMML (`a14:m`)	Required for `rk-academic-visualize` math injection (`inject-math.py`)
LaTeX / TikZ	TeX Live 2022+ / MacTeX	Compiles `.tex` TikZ source to vector PDF	Required for local TikZ rendering, optional for core SOPs
Figure renderer & image viewer	Format-dependent	Rasterize PDF; render Mermaid/SVG, including sanitized selected SVG assets; open images for inspection	Required for verified figures and SVG-to-PNG assets; optional for core SOPs
Git	2.30+	Workspace and version tracking	Required
PowerPoint / Office	2010+	Viewing and natively editing rendered OMML equations	Optional (equations remain editable in modern Office/Word/PowerPoint)

Use the figure's native renderer, then open the rendered raster image and inspect spacing, connectors, labels, container bounds, and margins. If rendering or image inspection is unavailable, report the figure as unverified. Portable commands and layout thresholds live in the illustration guide. See the rendered two-pass figure and native-equation slide screenshot in the development log. Click a thumbnail to open a full image preview inside the page.

For new slide decks, run python3 skills/rk-academic-visualize/scripts/check-visual-plan.py slide-plan.json --require-visual alongside the existing deck checker. If the deck uses external assets, pass visual-ledger.json before --require-visual; the visual asset guide covers license checks, visible credit, and safe SVG-to-PNG conversion. The visual linter helps review text density to optimize visual clarity, while accepting justified text-only slides.

Environment Verification Commands

# Verify Python version and suite integrity
python3 --version
python3 scripts/check-suite.py

# Verify Pandoc availability for math injection
pandoc --version

# Run standalone slide and math test suites
python3 -m unittest discover skills/rk-academic-visualize/scripts/tests

Installation & Quick Start

1. Using the Skills CLI (Recommended)

Install the entire suite directly into your active project or agent environment via the skills CLI:

npx skills add vinhnt21/research-kit

[!NOTE] Antigravity Scope Note:
Workspace Scope (Recommended): Running npx skills add vinhnt21/research-kit inside your project root installs skills into .agents/skills/, which Antigravity automatically discovers for that workspace.
Global Scope: Do not use npx skills add -g for Antigravity. npx skills places global skills in ~/.agents/skills/, but Google Antigravity IDE strictly discovers machine-global skills from ~/.gemini/config/skills/ (and CLI from ~/.gemini/antigravity/skills/). For global Antigravity installation, use GitHub CLI:
gh skill install vinhnt21/research-kit --dir ~/.gemini/config/skills --all
or manually copy: cp -r skills/rk-* ~/.gemini/config/skills/.

2. Using GitHub CLI (v2.90.0+)

Target specific AI agent environments with native GitHub CLI commands:

# Target Cursor
gh skill install vinhnt21/research-kit --agent cursor

# Or target other environments:
# --agent claude-code
# --agent codex
# --agent antigravity

# Antigravity IDE Global Scope:
gh skill install vinhnt21/research-kit --dir ~/.gemini/config/skills --all

Updating installed skills

Both installers can refresh already-installed skills after their source changes. Use the same tool you originally used to install them:

# Skills CLI: check and update installed skills in active project (Workspace scope)
npx skills update

# GitHub CLI: preview updates, then apply them interactively
gh skill update --dry-run
gh skill update

# Google Antigravity / Gemini (Update Machine-Global scope from local clone):
cd path/to/research-kit && git pull && cp -r skills/rk-* ~/.gemini/config/skills/

The Skills CLI also accepts npx skills upgrade as an alias for update; GitHub CLI uses gh skill update. For unattended updates, use npx skills update -y or gh skill update --all. Both CLIs also accept skill names to update only selected skills, for example npx skills update rk-ai or gh skill update rk-ai. A newly added skill is not an update to an existing installation; rerun the corresponding install command to add it. See the official Skills CLI update guide and gh skill update reference.

3. Agent-Assisted Installation (via ZIP)

Download the repository ZIP: research-kit-main.zip.

Prompt your AI agent (Cursor, Claude Code, Codex, Antigravity) to install automatically:

Extract the downloaded research-kit-main.zip and copy all folders from skills/rk-* into your skills directory (~/.cursor/skills/, ~/.claude/skills/, ~/.codex/skills/, or ~/.agents/skills/).

4. Manual Installation

Copy the skills/rk-* folders directly into your target agent's skill directory:

Cursor: ~/.cursor/skills/
Claude Code: ~/.claude/skills/
Codex: ${CODEX_HOME:-$HOME/.codex}/skills/
Antigravity (Global IDE): ~/.gemini/config/skills/ (or ~/.gemini/antigravity/skills/ for CLI)
Antigravity (Workspace): .agents/skills/ (inside project root)
Universal Agents: ~/.agents/skills/

5. Verification & Test Suite

Verify skill integrity, frontmatter bounds, and link graphs:

python3 scripts/check-suite.py

Regenerate public documentation figures (requires XeLaTeX and Ghostscript):

python3 scripts/render-doc-figures.py

Documentation

Public guides live in documents/:

Guide	Description
`documents/guide.md`	Full operational handbook (English): install/update/uninstall, managing multiple papers in one repo, per-skill reference, end-to-end playbook, decision matrix, and anti-patterns
`documents/guide.vi.md`	Same handbook in Vietnamese

Browse the interactive version on the site: Docs.

Comparison & Key Advantages

Scientific inquiry using AI agents currently suffers from severe standing context bloat, routing hallucinations, and brittle framework dependencies. Research Kit was engineered specifically to solve the architectural bottlenecks found in earlier libraries like Scientific Agent Skills (Kassis et al., 2026) and Science Superpowers (K-Dense-AI).

Direct Comparison: Scientific Agent Skills (v2.65.0) & Science Superpowers vs. Research Kit

Dimension	Scientific Agent Skills (v2.65.0, 163 Skills) & Prior Art	Research Kit (10 Streamlined Skills)	Practical Research Advantage
Catalog Architecture	Scientific Agent Skills: 163 fragmented, overlapping mini-skills across 16 domains	10 end-to-end lifecycle skills	Streamlined & convenient; eliminates tool choice overload
Research Focus	Scientific Agent Skills: Agent gets lost in tool search & install loops across 163 candidates	Deterministic stage-by-stage routing	100% focused on core scientific discovery
Standing Token Overhead	Scientific Agent Skills: 14,246 tokens (7.12% of context window) before reading files	~1,000 tokens (<0.50% of context window)	Frees >99.5% of context for raw data, papers, and deep reasoning
Operational Reliability	Scientific Agent Skills & Superpowers: 105 external scripts & brittle session-start hooks	Markdown SOPs + focused CLI checks and native figure renderers	Inspectable tools and explicit verification gates; core SOPs add no required dependencies
Framework & Multi-Paper Repo	Science Superpowers: Forces `docs/science-superpowers/` and git freeze; lacks multi-paper-in-one-repo support	Active Paper Declaration, strict evidence boundaries, and flexible layouts	Guaranteed data integrity across manuscripts in the same repo with zero forced lock-in

Key Advantages & Architectural Principles

1. Standing Context Window Economy (<0.50% Footprint)

Under the Agent Skills specification, an agent platform preloads each skill's name and description into the persistent system prompt at session initialization.

Scientific Agent Skills (v2.65.0) Overhead: 163 resident descriptions consume 14,246 tokens—over 7.12% of a standard 200,000-token window before the agent reads a single word of your paper or data (Kassis et al., 2026).
Research Kit Economy: Research Kit provides exactly 10 neighbor-aware skills. The combined descriptions total 2,335 characters and 327 words, consuming <0.50% (<1,000 tokens).
Practical Benefit: Research Kit conserves over 13,000 tokens on every prompt turn, preserving critical reasoning capacity for raw experimental data, extensive literature synthesis, and complex paper drafting.

2. Deterministic Stage-by-Stage Routing (Neighbor-Aware)

Each skill owns exactly one well-defined research stage and explicitly references its neighboring steps. An agent working on literature remains within rk-survey. A locked methodology belongs to rk-method. A claim lacking a fresh execution checkpoint stays in rk-data. Routing happens naturally and conveniently without hallucinated tool searches across 160+ options.

3. Multiple Papers in One Repo & Evidence Boundaries

In empirical research, a single repository frequently hosts multiple related ideas, hypotheses, or manuscripts concurrently. Without rigorous boundaries, AI agents easily trigger cross-contamination: reusing uncommitted exploratory numbers as baselines, or importing unverified claims across sister drafts.

my-research-project/           # Overall research topic (1 Repository / Root Workspace)
├── AGENTS.md                  # Global context, conventions, and lab guidelines
├── literature/                # Shared reference library (Read-Only)
│   ├── references.bib         # Global BibTeX citation repository
│   └── pdfs/                  # Reference papers, PDFs, and preprints
│
├── papers/                    # Partitioned workspaces for individual ideas/papers
│   ├── 2026-quantum-routing/  # [Idea A / Paper 1] - Separate workspace in same repo
│   │   ├── AGENTS.md          # Active Paper declaration: scope, questions, hypotheses
│   │   ├── src/               # Dedicated experimental code for Paper 1
│   │   ├── data/              # Dedicated data & independent rerun verification logs
│   │   ├── figures/           # Rendered publication figures for Paper 1
│   │   └── manuscript/        # Manuscript source files (LaTeX / Markdown)
│   │
│   └── 2026-repeater-sched/   # [Idea B / Paper 2] - Separate workspace in same repo
│       ├── AGENTS.md          # Active Paper declaration: scope, questions, hypotheses
│       ├── src/               # Dedicated experimental code for Paper 2
│       ├── data/              # Dedicated data & rerun logs for Paper 2
│       ├── figures/           # Rendered publication figures for Paper 2
│       └── manuscript/        # Manuscript source files for Paper 2
│
└── shared/                    # (Optional) Shared utility scripts or verified libraries

Target Paper Declaration (Active Paper Resolution): The agent must identify and declare the active paper from the prompt, cwd, or local AGENTS.md before reading files or running experiments. If ambiguous, the agent must ask for clarification.
Strict Evidence Boundary: A sibling draft, sister experiment folder, uncommitted script, or unpublished result from another paper is never valid evidence, baseline data, or citation material for the active paper.
Audited Read-Only Literature Shelf: Centralized literature (literature/) is shared for universal reference, but every cited reference must be independently audited against the active paper's specific claims.
Zero Dogmatic Lock-In: Labs may name directories papers/, studies/, ideas/, or manuscripts/. Research Kit requires only the core operational invariant: each paper is a separate workspace in the same repo with an unambiguous Active Paper declaration.

4. Pure Procedural Rigor (Zero Framework Lock-In)

No Forced Directory Trees: Unlike Science Superpowers which mandates docs/science-superpowers/, session-start git hooks, and disruptive git freeze commits, Research Kit leaves directory structure and git control 100% to the project.
Pre-Registration as a Tool, Not a Dogma: Pre-registration is supported for prospective confirmatory trials where appropriate, without blocking mathematical derivations or exploratory prototyping.
Zero Proprietary Harness Dependencies: Operates independently of any single agent UI, runtime harness, or proprietary subagent protocol.

5. Live Documentation over Static Copies

Scientific libraries (such as Qiskit, QuTiP, PennyLane, PyTorch, and scikit-learn) evolve rapidly. Rather than distributing static, outdated copies of library manuals, skills instruct agents to query the official, live documentation of the packages actively installed in the environment.

Companion References & Field Templates

Each skill ships with localized procedural guidelines (references/) and production field templates (assets/). Through progressive disclosure, agents only load these operational scaffolds when a specific skill is actively triggered:

Skill	Companion Reference Documents (`references/`) & Field Templates (`assets/`)
`rk-survey`	`references/search-boundary.md`, `references/citation-check.md`, `references/evidence-map.mdassets/search-record.md`, `assets/citation-checklist.md`
`rk-idea`	`references/framing.md`, `references/hypothesis-quality.md`, `references/rivals-and-falsification.md`, `references/biases-and-fallacies.mdassets/hypothesis-record.md`, `assets/rival-matrix.md`
`rk-method`	`references/design-choice.md`, `references/power-and-precision.md`, `references/feasibility-and-freeze.mdassets/method-plan.md`
`rk-data`	`references/inspect.md`, `references/test-choice.md`, `references/figures.md`, `references/anomalies-and-rerun.mdassets/analysis-record.md`
`rk-write`	`references/claim-evidence.md`, `references/section-roles.md`, `references/paragraph-flow.md`, `references/review-and-response.mdassets/claim-evidence.md`
`rk-report`	`references/report-gate.mdassets/report-record.md`
`rk-quantum`	`references/model-checks.md`, `references/execution-boundary.mdassets/quantum-run-record.md`
`rk-ai`	`references/leakage-and-splits.md`, `references/evaluation.mdassets/ml-eval-record.md`

(Note: rk-quantum-network operates as a self-contained skill specification. skills/rk-academic-visualize packages scientific visualization guidelines with the author's verified presentation templates and deck verification test suite).

References & Prior Art

Research Kit synthesizes proven scientific methodologies while eliminating framework bloat. All rewritten skills contain fresh procedural text licensed under Apache-2.0. Upstream licenses remain with their original authors.

Prior Art / Source	Citation / Version	Bottleneck in Prior Work	Principle Retained in Research Kit	How Research Kit Resolves It
Scientific Agent Skills	`v2.65.0`arXiv:2609.00065	163 fragmented skills consuming 14,246 standing tokens (7.12% of context); 105 external scripts; 29 credentialed env vars; context overflow in 29/46 workflows.	Clear functional decomposition across research concerns: literature, methodology, empirical data, writing, quantum, and ML eval.	Condensed into 10 neighbor-aware skills (<1,000 tokens, <0.50% footprint); standalone CLI utilities (OMML/OOXML, visual-plan lint, SVG sanitization); 0 credentialed APIs; 100% focused on research workflows.
Science Superpowers	Commit `0374bdf`	Dogmatic "Iron Law" of mandatory pre-registration for everything; automated git freeze scripts; harness-specific startup hooks; forced `docs/science-superpowers/` directory layout.	Rigorous sequential quality gates: question framing, search boundary, protocol design, anomaly investigation, and rerun check before claiming.	Preserves all empirical quality gates without dogma: pre-registration is a method choice for confirmatory trials; 0 forced hooks/commits; project-owned directory layouts.
Research Paper Writing Skills	Commit `77e7c2c`	Single skill tightly coupled to a rigid ML/CV/NLP manuscript outline; lacks domain-specific experimental checks.	Structural writing discipline: single core idea per paragraph, strict 100% claim-to-evidence alignment, and structured rebuttal workflows.	Decoupled into `rk-write` for universal scientific manuscript drafting; domain-specific checks are delegated to modular extensions (`rk-quantum`, `rk-ai`, etc.).
NVIDIA Skills	Commit `d8519c5`	Large, vendor-specific product catalog with static CLI instructions and risk of accidental spend on cloud QPU backends.	Strict execution boundary discipline: local simulation is default; physical hardware/QPU execution requires explicit authorization.	Integrated safe local simulation defaults into `rk-quantum` with live documentation queries; physical QPU spend strictly requires explicit user approval.
Agent Skills Specification	Spec (1 Oct 2026)	Uncurated skill catalogs easily overwhelm the client's standing context window during session initialization.	Standardized directory structure: single `SKILL.md` per skill, lightweight `name` and `description` for preloading, dynamic on-demand asset loading.	Strict 10-skill ceiling with lightweight metadata (~100 tokens per skill), achieving full spec compliance under <0.50% standing context.