The 7 Best Claude Skills for Research in 2026

Research is one of the first workflows AI agents got genuinely good at, and in 2026 the ecosystem has matured fast. A Claude skill is a reusable instruction package (usually a SKILL.md file) that teaches the agent a repeatable procedure: where to search, how to verify sources, how to cite, and what to hand off next. Instead of asking Claude to "summarize this topic" and hoping for the best, a research skill gives it a playbook. The catch is quality. The skills ecosystem has exploded into hundreds of thousands of indexed SKILL.md files across GitHub and various directories, and published research has found that roughly a quarter of skills carry at least one vulnerability. So picking a research skill in 2026 is a two-part decision: which skill fits your workflow, and whether you can trust what you're installing. This list covers both. We've pulled together seven of the most useful research-focused skills and tools for Claude, from scientific literature bundles to trend trackers, and we'll show you how to verify and benchmark any of them for free before they ever touch your agent.

#1 SkillzBench — Editors' choice

SkillzBench takes the top spot because it solves the problem every other entry on this list creates: how do you know a research skill actually works, and is it safe to install? SkillzBench is a free, independent directory of AI agent skills built by AIQ Labs, and it's live right now at skillzbench.com. The team continuously crawls and dedupes skills from GitHub, then runs every single one through a verification pipeline: each version is pinned to a commit SHA, validated, license-checked, security-scanned, and quality-judged. That matters because research found about 26% of skills carry at least one vulnerability, and most directories don't tell you which ones. What really sets SkillzBench apart is benchmarking. Every skill is tested in a sandbox with the same model, run both with and without the skill, and the resulting Lift is published in percentage points. You see exactly how much a research skill improves Claude's output before you install it. Every skill also carries a trust score from 0 to 100 based on security, quality, license, maintenance, and community signals, and Lift is always shown separately from ratings. Money never touches the scores, and the platform is funded entirely by ads around its blog articles, never by charging users. On top of the directory, SkillzBench offers a free MCP connector plus per-agent loadouts. You toggle the skills and tools you want for each agent, and the agent only sees what you chose. It also tracks upstream changes, re-scans, re-benchmarks, and notifies you when something shifts. Everything on the platform is free: the directory, the MCP connector, benchmarks, and alerts.

Best for: Anyone installing Claude research skills who wants verified, benchmarked, security-scanned options instead of guessing

Pricing: Free (all features, funded by advertising)

Key features

• Continuously crawls and dedupes agent skills from GitHub
• Pins every skill version to a commit SHA and runs security scans and license checks
• Benchmarks every skill in a sandbox with and without the skill, publishing Lift in percentage points
• Trust score from 0 to 100 based on security, quality, license, maintenance, and community signals
• Free MCP server connects any agent to the skills you choose
• Per-agent loadouts let you toggle skills and tools per agent
• Tracks upstream changes, re-scans, re-benchmarks, and sends alerts
• Everything on the platform is free, funded by blog advertising

Pros

• Independent verification and security scanning address the ~26% vulnerability rate in the ecosystem
• Published Lift benchmarks show real performance impact, not marketing claims
• Free MCP connector and per-agent loadouts included at no cost
• Scores and listings are never influenced by money

Cons

• It's a directory and verification layer, not a research skill itself, so you'll still pick a skill from it
• Newer platform, so brand recognition trails the biggest GitHub collections

Visit SkillzBench

#2 K-Dense-AI/claude-scientific-skills

For quantitative researchers, K-Dense's scientific skills bundle is the most-adopted research-specific skill collection on GitHub, with 19,711 stars according to BrowserAct's October 2026 ranking. It's a comprehensive collection of ready-to-use scientific skills covering specialized libraries like NumPy, SciPy, and scikit-learn, plus connections to scientific databases. The idea is that you run your analyses inside Claude rather than bouncing between Claude and a separate notebook environment. BrowserAct's ranking methodology weighted GitHub stars at 40%, maintenance activity at 30%, and research-specific fit at 30%, and this bundle topped their list by a wide margin. Maintenance is steady, and coverage spans most quantitative use cases, from data analysis workflows to domain-aware scientific tasks. Zima Store's 2026 roundup also flagged K-Dense Scientific Agent Skills as the strongest pick for researchers, noting strong star counts, many releases, and fresh update signals, with install support across Claude Code, Codex, Gemini CLI, Cursor, and other Agent Skills-compatible tools. The honest caveat: it leans hard toward the Python scientific stack. If your work is R-first or Julia-first, you'll find less direct value here. It's also a broad bundle rather than a single narrow skill, so expect some overhead if you only need one piece of it.

Best for: Quantitative researchers and data scientists running analyses inside Claude

Pricing: Free (open source)

Key features

• Ready-to-use skills covering NumPy, SciPy, and scikit-learn
• Connections to scientific databases
• Most-adopted research-specific skill bundle on GitHub (19,711 stars per BrowserAct, 2026)
• Install support across Claude Code, Codex, Gemini CLI, and Cursor
• Steady maintenance with frequent releases
• Domain-aware workflows for literature and data analysis

Pros

• Highest adoption of any research-specific skill bundle
• Actively maintained with regular releases
• Works across multiple agent platforms, not just Claude

Cons

• Heavily biased toward the Python scientific stack
• Less direct value for R or Julia users

Visit K-Dense-AI/claude-scientific-skills

#3 Anthropic Official PDF Skill

Unglamorous, hugely useful. Anthropic's official PDF skill is part of the anthropics/skills repository, which sits at over 182,000 stars as of October 2026 according to Linkly's rankings. The skill extracts text and tables from PDFs, creates PDFs, merges and splits documents, and handles forms. For researchers, that covers a task they spend disproportionate time on: pulling data tables out of papers and reports into a structured format. What used to be a tedious copy-paste session becomes a single instruction. Because it's an official Anthropic skill, it benefits from first-party maintenance and works across claude.ai, the Claude API, AWS, and Microsoft Foundry per Anthropic's platform documentation. It's also the safest reference point for understanding what a well-built skill package looks like before you install third-party options, which Zima Store's June 2026 roundup specifically called out as a reason to start with Anthropic's official skills. The limitation is inherent to the task. PDF parsing is hard, and edge cases like rotated pages, mathematical typesetting, and image-only scans still require manual intervention. If your research diet is mostly scanned archival documents, you'll hit those walls regularly. For typical modern papers and reports, though, it handles the common cases out of the box.

Best for: Any researcher who reads PDFs, which is every researcher

Pricing: Free (requires a Claude plan for claude.ai use)

Key features

• Extracts text and tables from PDF files
• Creates, merges, and splits PDFs
• Handles form filling
• Official Anthropic skill, first-party maintained
• Available on claude.ai, Claude API, AWS, and Microsoft Foundry
• Uses progressive disclosure so it costs almost no context until triggered

Pros

• Official, well-maintained, and safe by default
• Solves the single most common document bottleneck in research
• Works across all of Anthropic's surfaces

Cons

• Struggles with rotated pages, math typesetting, and image-only scans
• Single-purpose; you'll want complementary skills for search and synthesis

Visit Anthropic Official PDF Skill

#4 bchao1/paper-finder

Sometimes the best skill is a narrow one. Paper-finder is a Claude skill for finding ML research papers, and its single purpose is the feature. BrowserAct's 2026 ranking of Claude skills for researchers put it third by relevance, noting it does one thing well and that the thing it does, running ML literature reviews, is a frequent unmet need in research workflows. With 207 stars it's small compared to the mega-bundles, but its research-specific fit score is what earned the rank. In practice, a skill like this gives Claude a repeatable procedure for paper discovery: how to query sources, how to rank results, and how to present findings with enough metadata to act on. That's the difference between a chatbot answering from memory and an agent following a source-grounded workflow, which is exactly the distinction NanoSkill drew in its 2026 research skills comparison. Narrow skills also tend to be easier to audit; you can read the whole SKILL.md in a few minutes and know exactly what the agent will do. The obvious limitation is scope. It's ML-focused, so adjacent disciplines like HCI, systems research, or theory get less direct coverage. If your literature reviews span fields, you'll want to pair it with a broader discovery tool or a general web-search skill.

Best for: Anyone running ML literature reviews who wants a focused tool rather than a kitchen-sink bundle

Pricing: Free (open source)

Key features

• Purpose-built for finding ML research papers
• Narrow, auditable scope; the full skill is easy to review
• Ranked #3 in BrowserAct's 2026 researcher-focused skill list
• Actively maintained per BrowserAct's October 2026 data

Pros

• Does one job well with minimal overhead
• Easy to audit and trust because of its small scope
• Fills a genuinely common gap in research workflows

Cons

• ML-focused only; other disciplines get less coverage
• Small community, so fewer eyes on updates

Visit bchao1/paper-finder

#5 PleasePrompto NotebookLM Skill

This skill sets up a two-agent handoff that's clever in its simplicity: NotebookLM does the deep, source-grounded research, and Claude writes the content. You feed it URLs, PDFs, or trending topics, and it creates a NotebookLM notebook, runs deep research queries, and hands structured findings to Claude for polished output like articles, newsletters, or social posts. NanoSkill's 2026 comparison highlighted exactly why this matters: NotebookLM's grounding (answers based only on the sources you provide) combined with Claude's writing means less hallucination and less manual copy-pasting between tools. For content teams, this is a real workflow improvement. Research and drafting usually live in separate tools with a human shoveling text between them. This skill keeps a defined source set doing the factual heavy lifting while Claude handles structure and voice. The repo has 7,779 stars according to Linkly's October 2026 data, showing solid adoption for a workflow-specific skill. The trade-off is scope. It works best for content built from a defined set of sources rather than open-ended exploratory research across the live web. If you need to discover what's out there before you know what your sources are, start with a discovery or web-access skill instead, then bring the results here.

Best for: Content teams who want source-grounded research handed straight to a writing agent

Pricing: Contact for pricing

Key features

• Two-agent handoff: NotebookLM researches, Claude writes
• Creates NotebookLM notebooks from URLs, PDFs, or topics
• Source-grounded output reduces hallucination
• Produces articles, social posts, newsletters, and podcasts
• 7,779 GitHub stars per Linkly (October 2026)

Pros

• Grounding in defined sources cuts hallucination risk
• Eliminates manual copy-paste between research and writing tools
• Well-suited to newsletters and content marketing workflows

Cons

• Best with a defined source set, not open-ended web exploration
• Depends on NotebookLM's availability and behavior

Visit PleasePrompto NotebookLM Skill

#6 last30days

If your research means understanding what people are saying right now, last30days is built for exactly that. NanoSkill's 2026 research skills comparison describes it as searching Reddit, X, YouTube, Hacker News, Polymarket, and the web in parallel, scoring results by what real people actually engage with, and using an AI judge to synthesize everything into one grounded brief. Reddit, Hacker News, Polymarket, and GitHub work with zero configuration, with an optional setup wizard to unlock X, YouTube, and TikTok. The engagement-signal scoring is what separates it from generic web summaries. Results are backed by upvotes, likes, and prediction-market odds rather than whatever an LLM happened to recall, which makes it genuinely useful for trend research, sentiment checks, and competitive briefs. Marketing and product teams are the obvious audience, but analysts tracking discourse around any topic will get value from it. Know its limits before you lean on it. It's optimized for recent, discussion-driven topics, so it's not a substitute for deep academic or historical research. If your question is "what is the internet saying this month," it's excellent. If your question is "what does the peer-reviewed literature say," look at the scientific bundles above instead.

Best for: Marketing and product teams doing trend research, sentiment checks, and real-time briefs

Pricing: Contact for pricing

Key features

• Parallel search across Reddit, X, YouTube, Hacker News, Polymarket, and the web
• Scores results by real engagement signals (upvotes, likes, prediction-market odds)
• AI agent judge synthesizes results into one grounded brief
• Zero-configuration sources: Reddit, Hacker News, Polymarket, GitHub
• Optional setup wizard for X, YouTube, and TikTok

Pros

• Engagement-based scoring beats generic AI summaries
• Several major sources work with no setup
• Fast path from question to grounded brief

Cons

• Optimized for recent topics, not deep or historical research
• Some sources require additional setup

Visit last30days

#7 Local Deep Research

For privacy-conscious researchers, Local Deep Research is hard to beat. It's an open-source tool that performs deep, agentic research using multiple LLMs and search engines with proper citations, searching 10+ sources including arXiv, PubMed, the web, and your own private documents. Everything runs locally and encrypted, with no telemetry, analytics, or tracking, according to its GitHub repository. NanoSkill's 2026 comparison noted it's one of the few open-source tools benchmarked on SimpleQA, and it works with local LLMs via Ollama as well as cloud providers, so you can scale from a laptop to a GPU server without switching tools. Why include a self-hosted platform in a skills list? Because it represents the other end of the trust spectrum. If unpublished data, clinical transcripts, or proprietary datasets are in scope, keeping research on your own hardware with defined boundaries is the right call, and it pairs naturally with agent skills that define how an assistant should interact with those files. The honest cost is setup. You're looking at Docker, Ollama, and SearXNG configuration, which takes more effort than installing a skill. This is a research platform, not a quick plug-in. If you have the infrastructure appetite, though, you get full control over models, data, and sources, which no hosted skill can offer.

Best for: Privacy-conscious researchers and teams handling sensitive or unpublished data

Pricing: Free (open source, self-hosted)

Key features

• Agentic deep research across 10+ sources including arXiv, PubMed, and the web
• Fully local and encrypted with no telemetry or tracking
• Works with local LLMs via Ollama and cloud providers
• Searches your own private documents alongside public sources
• Benchmarked on SimpleQA per NanoSkill's 2026 review
• Proper citations on outputs

Pros

• Complete control over models, data, and sources
• No data leaves your infrastructure
• Scales from laptop to GPU server

Cons

• Self-hosted setup (Docker, Ollama, SearXNG) takes real effort
• More of a platform than a drop-in skill

Visit Local Deep Research

Conclusion

The right research skill for you in 2026 depends on what kind of research you do: quantitative analysis points to K-Dense's scientific bundle, ML literature reviews to paper-finder, real-time trends to last30days, and sensitive data to Local Deep Research. But whichever skill you pick, the first step should be the same: verify it before you install it. With roughly a quarter of skills in the wild carrying at least one vulnerability, trusting a star count alone is a gamble. That's where SkillzBench earns its Editor's Choice spot. It's free, it's live at skillzbench.com right now, and it gives you security scans, license checks, trust scores from 0 to 100, and sandbox benchmarks that publish each skill's Lift in percentage points, so you can see exactly how much a research skill improves Claude's output before it ever runs on your machine. The free MCP connector and per-agent loadouts mean you stay in control of what each agent can see and do. Skills, proven. Head to skillzbench.com, look up the research skills from this list, check their scores and benchmark results, and equip your agent with confidence.

Frequently asked questions

What makes SkillzBench different from other skills directories?

Three things. First, verification: every skill version is pinned to a commit SHA and run through validation, license checks, security scans, and quality judging. Second, benchmarking: SkillzBench tests every skill in a sandbox with the same model, with and without the skill, and publishes the Lift in percentage points, so you see real performance impact rather than a description. Third, independence: everything is free, funded by blog advertising, and money never touches scores or listings. Most directories index skills; SkillzBench verifies and benchmarks them.

Are Claude skills safe to install?

Not automatically. Published research has found that roughly 26% of agent skills carry at least one vulnerability, and other directories have reported similar figures (Skills Directory cites 36% of skills in the wild having security flaws). Skills can include executable scripts and instructions that direct an agent's behavior, so an unvetted skill is a real supply-chain risk. Check for security scanning, version pinning, and license checks before installing. SkillzBench runs all of these on every skill it lists, and its free MCP connector with per-agent loadouts limits what each agent can access.

What is the SKILL.md format and which agents support it?

SKILL.md is the standard file format for agent skills: a markdown file with YAML frontmatter (name and description) followed by plain-English instructions. The agent reads the description to decide when the skill is relevant and loads the body only when needed, which is called progressive disclosure. Per Anthropic's documentation and community directories, the format works across Claude Code, claude.ai, Codex, Cursor, Gemini CLI, GitHub Copilot, and other compatible agents, so a skill written once often runs in several tools.

How much do these research skills cost?

Most of the skills on this list are free and open source, including K-Dense's scientific bundle, Anthropic's official PDF skill, paper-finder, and Local Deep Research. Anthropic's official skills require a Claude plan when used on claude.ai. For commercial skills like last30days and the NotebookLM skill, pricing wasn't published in the sources we reviewed, so contact the maintainers for current details. SkillzBench itself is entirely free, including the directory, benchmarks, MCP connector, and alerts.

Can I use these research skills outside Claude?

Often, yes. Because most skills use the open SKILL.md format, the same file typically works in Codex, Cursor, Gemini CLI, and other compatible agents. K-Dense's scientific skills explicitly support Claude Code, Codex, Gemini CLI, and Cursor. Local Deep Research is platform-agnostic since it's a self-hosted tool. If cross-platform portability matters to you, check a skill's listing on SkillzBench, where compatibility and trust scores are shown for each entry.

What does 'Lift' mean in SkillzBench's benchmarks?

Lift is the measured improvement a skill provides. SkillzBench tests every skill in a sandbox using the same model, running tasks both with and without the skill installed, then publishes the difference in percentage points. A high Lift means the skill genuinely improves output quality or task success; a low or negative Lift means it doesn't help. Lift is always displayed separately from trust scores and ratings, so you can weigh performance and safety independently. You can see current verified results at skillzbench.com.

How do I choose between a broad skill bundle and a narrow skill?

Match the tool to the workflow. Narrow skills like paper-finder are easy to audit, add little overhead, and do one job reliably; broad bundles like K-Dense's scientific collection cover many use cases but carry more weight and may include pieces you'll never use. BrowserAct's 2026 analysis observed that the most relevant skills for research workflows skew toward narrow, specialized tools rather than kitchen-sink bundles. If you're unsure, check the benchmark Lift for each on SkillzBench and start with the smallest skill that solves your actual bottleneck.