Best Agent Skills in 2026: Picks by Task

Short answer: there is no single best agent skill. The right one depends on the job. Below are 21 skills for ten common tasks, from PDFs and spreadsheets to code review, CI/CD and research. For each one we give the publisher, licence, approximate size and what it does, all checked in Skill Store 33 on Oct 30, 2026.

Know this first: every pick below shows "Not yet verified" in the store, as do nearly all of its skills today. That means the skill was imported from its public source and its own tests have not yet run in the store's sandbox. It is not a warning, and it is not an endorsement. Read a skill's SKILL.md before you install it.

We update this page as the store's tests run. When a pick's state changes to Verified or Tests failing, we change it here.

How we picked

Publisher. Official vendor and project repositories came first: Anthropic, Vercel, OWASP, LangChain, Firecrawl, Nous Research, Cline. Then well-known community collections such as obra/superpowers. We list the publisher as the store shows it, which is the GitHub owner of the source repository.

Clear scope. One job per skill, and a description that says when to use it. An agent picks skills from their descriptions, so a vague one rarely fires.

Licence. Every pick has a permissive licence on its store page: MIT, Apache-2.0, ISC, CC0-1.0 or CC-BY-4.0. Search results also turn up third-party copies of Anthropic's source-available document skills. We left those out.

Size. The store shows roughly how many tokens the SKILL.md takes to load. Smaller is cheaper. Larger skills often carry scripts and reference files that do real work, so size is a trade-off, not a score.

1. PDF and document work

PDF, by nousresearch (MIT, about 3,100 tokens). From Nous Research's Hermes Agent repository. Creates PDFs from a JSON spec, builds and fills forms, extracts text and tables, merges, splits, stamps and encrypts, with OCR covered in a reference file. It is written for Hermes Agent and names that agent's tools, such as terminal, so other agents have to map those steps.

PDF, by appautomaton (MIT, about 1,000 tokens). Reads, creates, fills and transforms PDFs. It tells the agent to render the pages it changed and look at them before handing the file back, and it lists the mistakes that fail silently: two coordinate systems, rotated pages and missing fonts for non-Latin text.

2. Frontend design

Frontend Design, by cline (Apache-2.0, about 1,100 tokens). From Cline's skills repository. Builds web components, pages and apps with a deliberate visual direction, and steers away from generic AI-looking design.

Vercel React Best Practices, by vercel-labs (MIT, about 1,800 tokens). React and Next.js performance guidelines from Vercel Engineering. Use it when writing, reviewing or refactoring components, data fetching or bundles.

3. Code review

Requesting Code Review, by obra (MIT, about 740 tokens). From the obra/superpowers collection. When a task or feature is done, it sends a code-reviewer subagent to check the work against the plan before you merge.

Code Review Security, by owasp (CC-BY-4.0, about 550 tokens). From OWASP's secure-agent-playbook. Reviews a pull request by vulnerability class, mapped to the OWASP Top 10 and ASVS. Each finding gives the file and line, an attack scenario, a fix and a confidence rating.

4. Testing

Test Driven Development, by obra (MIT, about 2,400 tokens). Also from obra/superpowers. Has the agent write a failing test before any implementation code, for every feature or bug fix.

Webapp Testing, by anthropics (Apache-2.0, about 970 tokens). From Anthropic's skills repository. The agent writes Python Playwright scripts against your local app. A bundled with_server.py helper starts and stops your dev servers around the test.

Deep dive: our post on agent skills for end-to-end browser testing compares more options.

5. CI/CD

GitHub Actions, by tartinerlabs (MIT, about 1,400 tokens). Audits your workflows by default and writes new ones only when you ask. It pins every action to a full commit SHA and checks permissions, concurrency and caching.

Setting Up CI, by spencerpauly (CC0-1.0, about 580 tokens). A short recipe for a lint, type-check, test and build workflow, with optional matrix and deploy jobs. Its example pins actions by version tag, such as actions/checkout@v4, not by commit SHA.

Deep dive: how to pick a CI/CD skill for your agent.

6. Data analysis

Explore Data, by anthropics (Apache-2.0, about 2,900 tokens). From the data plugin in Anthropic's knowledge-work-plugins repository. Profiles a new table or file: its shape, null rates, column distributions and quality problems such as duplicates, before you decide what to analyse.

Exploratory Data Analysis, by aj-geddes (MIT, about 1,700 tokens). Finds patterns, distributions and relationships through charts and summary statistics, and turns them into hypotheses to test.

7. Writing and docs

Documentation, by anthropics (Apache-2.0, about 370 tokens). From the engineering plugin in knowledge-work-plugins. Short outlines for a README, API docs, a runbook, an architecture doc and an onboarding guide, plus five writing principles.

Technical Writer, by onewave-ai (MIT, about 1,200 tokens). Writes or restructures READMEs, tutorials, runbooks, release notes and similar docs, grounded in the actual code, with commands checked where possible.

8. Building an MCP server

Build MCP Server, by anthropics (Apache-2.0, about 3,000 tokens). From Anthropic's claude-plugins-official repository. It asks about your use case before writing code, then recommends a deployment model (remote HTTP, MCPB or local stdio), a tool design and a framework, either the TypeScript SDK or FastMCP.

MCP Builder, by shipshitdev (Apache-2.0, about 3,500 tokens). Wraps an external API as tools an agent can call, in Python or Node/TypeScript.

9. Spreadsheets

Xlsx, by nousresearch (MIT, about 2,400 tokens). Hermes Agent's Excel skill. Its openpyxl scripts build, read and edit workbooks, insert rows without breaking formula references, and recalculate through LibreOffice. Like the PDF skill above, it names Hermes Agent's tools.

Spreadsheets, by paulrberg (MIT, about 1,500 tokens). CSV, TSV and Excel files, with DuckDB or qsv queries over them. It covers exact decimal arithmetic and escaping against formula injection, which matters for financial data.

Gws Sheets, by googleworkspace (Apache-2.0, about 650 tokens). From the googleworkspace/cli repository. Reads and writes Google Sheets through the gws command. It needs the gws CLI and its shared auth skill.

10. Research

Web Research, by langchain-ai (MIT, about 820 tokens). From LangChain's deepagents repository. Writes a research plan, sends up to three subagents in parallel with 3 to 5 searches each, then writes a report that cites its sources. It uses deepagents tool names such as task and write_file.

Firecrawl Deep Research, by firecrawl (ISC, about 1,400 tokens). For a formal, cited report, not a quick lookup. It asks how long you want it to run, then searches and scrapes with Firecrawl at that depth. It needs Firecrawl.

Before you install any of these

A skill runs with your agent's permissions. Open the SKILL.md and every script it bundles, check the licence and the source repository on its store page, and pin the version you reviewed. Our checklist post walks through it in five steps.

FAQ

What is the best agent skill?

There isn't one. The right skill depends on the job and on your agent. Pick by publisher, scope, licence and size, and read the SKILL.md before you install it.

What does Not yet verified mean?

The skill was imported from its public source and its own tests have not yet run in the store's sandbox. The other states are Safety scanned (a safety scan only), Verified (its tests ran in a sandbox and passed) and Tests failing.

Why aren't Anthropic's own PDF and Excel skills here?

Anthropic's skills repository says its docx, pdf, pptx and xlsx skills are source-available, not open source. If you want them, install them from Anthropic, for example with /plugin install document-skills@anthropic-agent-skills in Claude Code after adding the anthropics/skills marketplace.

How do I install one of these skills?

Download the SKILL.md from the skill's store page into your agent's skills folder, or let your agent fetch it through the MCP server below. Several picks bundle scripts; for those, copy the whole folder from the source repository linked on the store page. The install guide lists the folder for each agent.

Let your agent search Skill Store 33

Sources

Skill Store 33 skill pages and records for all 21 picks (publisher, licence, token size, verification state, source repository), linked above (read Oct 30, 2026)

Anthropic, anthropics/skills README (licensing of the document skills, install commands), https://github.com/anthropics/skills (read Oct 30, 2026)

Source repositories of the picks: https://github.com/nousresearch/hermes-agent, https://github.com/appautomaton/document-skills, https://github.com/cline/skills, https://github.com/vercel-labs/agent-skills, https://github.com/obra/superpowers, https://github.com/owasp/secure-agent-playbook, https://github.com/tartinerlabs/skills, https://github.com/spencerpauly/awesome-cursor-skills

More source repositories: https://github.com/anthropics/knowledge-work-plugins, https://github.com/aj-geddes/useful-ai-prompts, https://github.com/onewave-ai/claude-skills, https://github.com/anthropics/claude-plugins-official, https://github.com/shipshitdev/skills, https://github.com/paulrberg/agent-skills, https://github.com/googleworkspace/cli, https://github.com/langchain-ai/deepagents, https://github.com/firecrawl/skills

This article was created with the help of AI (AI-Assisted). Each skill's publisher, licence, size and verification state was checked in Skill Store 33, and its contents read, on Oct 30, 2026. ReadyAI is accountable for the content.

Keep reading