AI outreach that sounds like you: a practical method

AI outreach sounds like AI when it could have been sent to anyone. Tone is the visible symptom, but the causes are usually further back: a thin brief, no real detail about the reader, and a model free to fill the gaps. Fix those and the voice is much easier.

The tells people notice

• Openers that praise without saying anything: “I was really impressed by your journey.”
• Stock phrases: “I hope this finds you well”, “quick question”, “I'd love to pick your brain”.
• The same shape every time: compliment, value statement, soft ask.
• Being too complete. People write short notes with one idea. Models write tidy paragraphs covering everything.
• A register that doesn't fit the place. A LinkedIn paragraph pasted into Reddit reads as an advert.

Step 1: Describe the voice from real writing

Don't describe yourself as “friendly but professional”. That applies to everyone. Give the model real samples: posts, emails you've sent, a bio you wrote yourself. Then have it describe what it sees in concrete terms, and correct it.

Twelve years running field crews before moving into software. Writes short, plain sentences. Asks one question at a time. Never uses exclamation marks. Would say “that's a pain”, not “that sounds challenging”.

Specifics like the last line do more than any adjective. If several people send from the same company, each one needs their own: a shared “brand voice” is how a whole team starts sounding like one bot.

Step 2: Set rules per platform

The same person writes differently in different places, and each platform has its own limits. A sensible starting set:

Step 3: Give it real proof, and forbid anything else

This is the step that protects you. Keep a short list of what you can truthfully claim: results, customers you may name, numbers with their source. Tell the model that's the only proof it may use. If nothing fits, it should say nothing, not improvise. An invented statistic in a sales message is worse than a dull sentence, and it's the kind of error a tired reviewer misses.

Step 4: Feed it the thread, marked as untrusted

A reply should depend on what the person said, so the model needs the whole conversation. But that text comes from a stranger, so mark it as data, not instructions. Otherwise a message that says “ignore the above” can steer your own tool.

Step 5: Edit, and use the edits

Read drafts aloud. Delete the first sentence and see if the message improves, because it often does. When you keep making the same change, put it in the voice description. A week of edits teaches the model more than a clever prompt.

In Programmable Social

Each sender gets a voice from their LinkedIn profile PDF, a bio or a writing sample, read on your own AI key, and you can edit it. Drafts follow per-platform rules, may only use the proof you've supplied, and treat captured messages as untrusted. A person approves before anything goes.