Best AI Agent Skills for Journalists in 2026

Newsroom budgets keep shrinking, deadlines keep tightening, and the average general-assignment reporter is now expected to file 12-15 stories a week. AI agent skills, the small, installable extensions that bolt onto Claude Code, are starting to absorb the parts of journalism that used to eat the most time: source gathering, transcription cleanup, PDF wrangling, and document export.

This guide walks through eight skills I've watched journalists install and actually keep. The list is ranked by how often the skill survives past the trial week. The throwaway score on a lot of "AI for journalists" tools is brutal, most never get used twice. These eight do.

Each one is open source, free to install, and runs locally inside Claude Code. No subscription on top of your existing Claude usage.

How to install any of these

Before the list, the install pattern:

# In Claude Code
/skill install <skill-name>

# Or by repo URL
/skill install github.com/<author>/<repo>

Skills go in ~/.claude/skills/ by default. To list what's installed: /skill list. To remove: /skill remove <name>.

A security note before installing anything: every skill is code that runs on your machine. Check the repo, scan the SKILL.md, and verify the author isn't a brand-new account. The ones in this list are from named contributors or established orgs, but make this check a habit.

1. Deep Research

Repo: github.com/various-authors/deep-research Install: /skill install deep-research License: MIT

This is the skill I've seen survive the longest in journalist setups. It takes a research question, breaks it into sub-queries, runs them in parallel, synthesizes a draft with citations, and flags any claim it couldn't verify.

The workflow journalists use: 1. Drop a story angle into Claude Code: "find the 5 most-cited studies on remote work productivity since 2024 that include sample size ≥1000." 2. Deep Research splits this into a search plan, fetches sources, cross-references findings. 3. Output includes a bibliography with URLs, publication dates, and confidence ratings.

What it doesn't do: replace the phone call. It surfaces sources you might have missed. Investigative work still needs the human conversation.

2. PDF Tools

Repo: github.com/anthropic-skills/pdf Install: /skill install pdf-tools License: Apache 2.0

Court documents, government reports, academic papers, leaked memos, journalism runs on PDFs, and the formatting is always painful. PDF Tools extracts text, tables, and embedded images, and handles OCR for scanned documents.

The use case that won me over: extracting tables from a 300-page state agency PDF. The skill pulled all 47 tables, kept the column headers, and exported clean CSVs. The same job in Adobe Acrobat took 4 hours the previous month.

For redacted documents, the OCR also gives you a heads-up where redactions live, which matters if you're reverse-engineering what was hidden.

Full review of the PDF Tools skill.

3. Firecrawl

Repo: github.com/mendableai/firecrawl Install: /skill install firecrawl-skill License: AGPL-3.0

When you need to scrape a site's structured data, court calendars, lobbying disclosure databases, public meeting transcripts. Firecrawl handles JavaScript-rendered pages and pagination that simple curl scripts choke on.

Journalist use case: monitoring a campaign finance database that updates weekly without an RSS feed. Firecrawl can crawl on a schedule and diff the output, so you get an alert when a new filing appears.

Note on terms of service: scraping public-record sites is legally settled in the U.S. (post Van Buren), but private platforms still have anti-scraping clauses. Use this for government and public-record sources, not for LinkedIn or Twitter scraping.

4. DOCX (Word Documents)

Repo: github.com/anthropic-skills/docx Install: /skill install docx-word License: Apache 2.0

Editors live in Word. They mark tracked changes, drop comments, and expect a clean .docx back. The DOCX skill reads existing Word files (including tracked changes and comments), writes new ones with proper formatting, and handles the round-trip without breaking the editor's markup.

The reason this matters: most copy-paste from Claude into Word loses formatting and forces you to re-create headings, indentation, and footnotes. The DOCX skill produces a Word file with all of that intact, which keeps the editor in their tool.

For newsrooms still using AP Style guides, you can also ask Claude to apply AP Style edits before exporting, handy for converting outside contributors' drafts.

5. Exa MCP Server

Repo: github.com/exa-labs/exa-mcp-server Install: /skill install exa-mcp-server License: MIT

Exa is a search index optimized for semantic queries, useful when Google's keyword matching doesn't find what you need. The MCP server lets Claude search Exa directly and use the results to write or fact-check.

Where it beats regular search: queries like "find local news stories about ICE raids at workplaces in 2025 in states without sanctuary policies." Keyword search will return everything ever written about ICE. Exa returns the specific intersection.

The free tier covers most journalist workloads. If you're running automated monitoring across many beats, you'll hit the paid tier eventually.

6. MCP Server YouTube Transcript

Repo: github.com/kimtaeyoon83/mcp-server-youtube-transcript Install: /skill install mcp-server-youtube-transcript License: MIT

For any story that involves a public video, press conferences, congressional hearings, viral creator content, this skill pulls the YouTube auto-captions or the official transcript, formats them with timestamps, and lets Claude analyze them.

Use case: a 3-hour Senate committee hearing. Asking Claude to "find every mention of 'inflation' and list the speaker, quote, and timestamp" produces a clean source list in under a minute.

Quality note: YouTube auto-captions are about 85-92% accurate for clear English speakers. For mumbled audio or strong accents, you'll still need to verify against the video. Don't quote from an auto-caption transcript without listening.

7. Firecrawl (alternative install)

There are two Firecrawl skill packages floating around the registry. The main one above is fine for most journalists. Power users sometimes prefer the community-maintained variant which adds custom output formats. Pick one and stick with it.

8. Claude SEO (yes, really)

Repo: github.com/various-authors/claude-seo Install: /skill install claude-seo License: MIT

Most journalists wave SEO off as somebody else's job. Then a piece they spent four weeks reporting gets 2,000 pageviews because the headline failed in search.

The Claude SEO skill audits a draft headline and standfirst against current ranking patterns, suggests three alternates, and flags missing schema markup the CMS should have added. For newsroom freelancers who own their bylines, this matters even more, the difference between a piece that ranks for 3 years and one that vanishes in 6 months is mostly title structure and entity coverage.

It is not about gaming Google. It is about not actively hiding your work from people searching for it.

What's missing from this list

I left out a few categories on purpose:

• AI transcription skills: the best transcription is still happening in Otter.ai, Trint, or Tactiq as standalone tools, not as Claude Code skills. The skills for this category aren't yet at parity.
• Generative image skills: ethical landmines for newsrooms. Don't install these for editorial work.
• Fact-checking-specific skills: none mature enough to recommend. Snopes, PolitiFact, and your own phone are still better than any automated fact-check skill on the market.

For the broader landscape, see our list of AI tools for journalists — covers transcription, monitoring, and visualization tools that sit outside Claude Code.

A short note on attribution and disclosure

Most major newsrooms (AP, Reuters, Bloomberg, the New York Times) have written disclosure policies on AI use. The trend across all of them: AI-assisted research and editing is fine if disclosed and verified; AI-generated original copy is not. These skills sit in the first category, they help with grunt work, not with original reporting.

If your newsroom hasn't published a policy yet, write one. Your readers will eventually ask.

FAQ

Are these skills safe to install on a newsroom-managed laptop? Most are. Check with your IT or security team before installing anything that runs scrapers or makes outbound API calls, some IT policies restrict outbound traffic from journalist laptops. The DOCX, PDF Tools, and YouTube Transcript skills are the safest from a security perspective.

Do I need Claude Pro or Claude Max? Claude Pro is enough for most journalist workloads. Max ($200/mo) makes sense if you're running heavy Deep Research workflows daily or processing long PDFs in batches.

Can these skills replace my newsroom's research desk? No. They make the research desk faster and let smaller teams cover more ground, but the human research librarian is still the highest-leverage person on a newsroom staff. If anything, AI raises the value of having one who knows what questions to ask the tools.

What about confidential source handling? None of these skills should touch material from a confidential source. Anything that goes into Claude is potentially logged. For protected-source material, stay on air-gapped local tools and never paste into a cloud LLM.

How often should I update installed skills? Check monthly. The skill ecosystem moves fast and security patches matter. /skill update <name> keeps each one current.