Create an outreach agent for your brand

Sets up everything outreach-day needs to run a brand's outreach: who to reach and where, what the brand may and may not say, and how it sounds. You read the brand's website, ask its owner what a website cannot tell you, and write the agent folder for them to review. Nothing touches the phone, and nothing is sent.

Inputs

Run it

Every command is node <outreach-day's folder>/scripts/outreach.ts <command>; run npm install --prefix <outreach-day's folder>/scripts once first. Below it is written outreach.

1. Read the brand. Fetch the site: the home page, pricing, the about page, the FAQ or help center, the terms. Note, with the page each came from: what the product is and is not, who it is for, prices and plans, any promotion, claims the site itself makes, and anything regulated (health, finance, legal, immigration, insurance, employment).

2. Ask the owner what the site cannot say, all at once, offering your best guess for each from what you read:

• Who exactly to reach: job titles, or the kind of person, and where they are (a country, schools, industries, cities).
• Which channels and accounts: LinkedIn, X, Reddit (which subreddits), Facebook (which groups the account has joined), Xiaohongshu. Each needs the brand's account signed in on the phone.
• How many a day: suggest modest numbers (LinkedIn 10-15 invites a day and under 100 a week; 4-10 comments a day per channel) and say why.
• What must never be said: advice the product does not give, claims that would need a disclaimer, competitors not to mention.
• Who to skip: competitors, firms or vendors selling to the same audience.
• Review or autopilot: whether writes wait for them or go out within the limits. Recommend review until a few runs have been checked.

3. Write the folder.

outreach new --dir <folder> --id <id> --site <site> --label "<who, for what>"

Add --records ~/.outreach/<id> when the folder will live in a repo others can read: the records hold people's names.

Then fill in every CHANGEME in the three files, from what you read and what they answered:

• agent.yaml: audience (titles, exclude_headlines, the rotation of schools, industries or cities), the channels they chose, with keywords people actually use when they need the product (their words, not the brand's), exclude_terms for vendors' words, the daily limits, order, offer (prices exactly as on the site) and the warmup message, which takes its prices from offer as {{min_price}}, never typed in.
• brand.md: what the product is and, more important, what it is not; the facts that may be stated, with where each came from; what must never be said; the safe redirect for anything you are unsure of. Nothing the site does not say: no invented numbers, outcomes or testimonials.
• voice.md: what a comment is for (usually: earn the profile visit, not sell), the angles per kind of post, who to skip, and each channel's rules.

Anything else the agent should read before it writes (a list of help articles to link, notes) goes in the folder as another .md file.

4. Check it.

outreach lint --agent <folder>

Fix what it lists until it says ready: placeholders left, a channel that plans no searches, a message that does not render.

5. Show the owner the three files and a sample of what the agent would write: two or three comments for posts you can imagine in their channels, and the warmup message (outreach message --agent <folder>). Change what they want changed. Then tell them how to run it: outreach-day with this folder, interactively first, and their own scheduler (cron, or their AI's scheduled tasks) once they trust it.

What comes back

The agent folder, ready to run, and a short summary: who it reaches, on which channels, how many a day, review or autopilot, and the commands to run it.

Notes

• outreach-day/example/ is a complete fictional agent to compare against.
• Regulated products need the strictest brand.md: say exactly what the product is not, and give the redirect for anything that sounds like advice.
• Never put the owner's API keys, passwords or private data in the folder.