Newsjack is 18 open-source PR skills you install into Claude once. By the end of this post you'll have them running in chat, a journalist list with verified email addresses, and a Claude Code agent that checks the news overnight and leaves a pitch-ready brief every morning at 8 a.m. Everything is public at newsjack.sh — skills, source code, updates.
This is the written version of my 27-minute install video; the canonical walkthrough has the embed and chapter timestamps if you'd rather watch.
Step 1 — Install the skills
Two options.
Option A: paste one prompt. Scroll down on newsjack.sh to the install instructions, copy the install prompt, paste it into any AI chat, and send. The agent walks itself through setup.
Option B: Claude's Customize tab.
- In Claude, open Customize in the left sidebar (this is where any Anthropic plugin installs).
- Plus sign → Create plugin → Add a marketplace.
- Choose Add from a repository — it asks for a GitHub repository or git URL.
- Copy the project URL from newsjack.sh, paste it in, hit Sync. Claude finds the project.
- Click the plus sign on the found project. You get a detail view of exactly what came in.
Verify the install by typing a forward slash in any new chat:
/
The slash menu lists every skill available to you, including everything Newsjack just brought in.
Step 2 — Run the chat skills
A good starting sequence is strategy → angles → headlines.
/pr-strategist takes a URL or a plain paragraph:
/pr-strategist
I'm an AI engineer. How can I do PR for myself?
It researches and comes back with your advantages and suggested positionings. While a skill loads, click it to read exactly what it's asking the model to do — every skill is a file in the open-source repo, and installing the plugin gives you an exact copy of what's online. No black boxes.
/angle-generator takes a news story, searches the web, and brainstorms ways in.
/headline-generator is worth running before you write anything — generate headlines first, then work backwards toward the story. One-line prompts get passable output; give it context and the headlines get meaningfully better.
Step 3 — The part where LLMs hit a wall
Run /find-journalists and you reach the step where Claude or ChatGPT normally breaks: models have no verified journalist contact data, and they can't look up someone's past articles or understand their beat. So they do what models do with missing data — hallucinate email addresses, or confidently return a journalist who isn't current.
The fix is a data source, not a better prompt. The connector is bundled in the plugin: back in Customize, under Newsjack, click Connect. It links a Medialyst project — full disclosure, that's my product; it's what gives the agent journalist enrichment, contact info, and news search.
Gotcha from the video: after connecting, refresh the page and start a new chat. Paste in your previous chat history and /find-journalists picks up exactly where you left off.
The result is a table of journalists. Scroll right for the column that matters: actual, verified email addresses. Then iterate conversationally:
find me 10 more journalists who cover model releases like GPT or Opus
Step 4 — 24/7 monitoring needs a runtime
Website chat can't run long enough to watch the news while you sleep. Your runtime options:
| Runtime | Notes |
|---|---|
| Claude Code | included in the $20 plan most people already pay for |
| Codex | OpenAI's equivalent |
| Hermes Agent | open source, free to install |
| OpenCloud | open source, runs on your machine |
This walkthrough uses Claude Code since most people already have it.
Step 5 — Set up the daily monitor
Download the Claude desktop app. Top left has two modes: chat (what we've been doing, in a different window) and code — the flagship mode, running on your local machine and considerably more powerful. Installing Newsjack here is one sentence:
install this for me
Then ask for the monitor:
set up daily monitoring for me
Add context: who you are, what stories you're looking for, when to run (mine: 8:00 a.m. daily). You don't have to front-load everything — it asks when it needs something.
Two settings that matter before you kick it off:
- Model: Opus. For knowledge-intensive work like PR, the quality gap between Opus and a lesser model is large.
- Thinking effort: high at minimum — higher if your usage limits have room.
It does the setup, then runs a live test against the day's news. Mine ran almost nine minutes — and that's the point. Claude Code is built for long runs (I've had it going for hours, even days), which is also what unlocks fan-out workflows like brainstorming ten ideas and sending them to separate agents to evaluate from different angles. Chat can't do that.
When the test finishes, it asks where the monitor should live — a Claude Code routine is what I'd suggest.
What you get back
show me the report
Three tiers, waiting every morning at 8:
- Pitch-ready stories matched to the persona you gave it, angles already attached (generated by the same angle skill from Step 2)
- Big stories worth looking into
- Related stories to watch
It scales per client — one monitor each — and it'll format a brief into a PDF if you ask. Every skill works identically in both places: type / in Claude Code and call the same angle generator, headline generator, or find-journalists you used in chat.
Rough edges, honestly
- The connector needs that refresh-and-new-chat dance after linking. It's in the video so you don't hit it blind.
- Mapping which skill runs where (chat vs. the four runtimes) turned out far more complicated than I expected — hence the table above.
- The newsworthiness-check skill sounds boring but is foundational: a common failure mode is that AI thinks everything is news, and pitching a non-story ends badly. Start there before you automate anything outbound.
Full walkthrough with the video and all 14 chapter timestamps: medialyst.ai/blog/newsjack-install-walkthrough
Top comments (0)