Don't let AI write your messages. Let it fill one line.

Three layers. Only one is risky.

Every message is built from three kinds of text. Keep them apart and you stay in control.

Aim for a message where you wrote nearly everything, spintax covers the connective words, and the AI line is one sentence out of five or six.

The syntax. Copy, paste, done.

This is how each layer looks inside a Vetta sequence step.

Try it yourself

Edit the message. Add your own {options|like this}. Hit Shuffle to see sends.

Variants must be swappable. If you can't replace one with another without changing the meaning, it isn't spintax. It's a different message.

Six rules for the AI line.

The AI variable is the only place a model touches your message. Keep it on a short leash.

• One sentence. 12 to 18 words on LinkedIn, up to 25 in email. Put the cap in the prompt.
• One job. Name the signal that surfaced the contact. No compliment, no pitch, no question. Your script has those.
• First-party facts only. Things the contact wrote, or their own company says about itself. Nothing they merely reacted to.
• Allowed to say nothing. Thin or old facts mean an empty line. Blank is safe. Wrong is not.
• The message survives without it. Delete the AI line and read the script. It should still work.
• You read it first. Preview at least 10 real contacts before anything sends.

Five accuracy gates. Pass all five or cut the fact.

Any specific fact in an AI line (a product, a number, an event, a person) has to clear every gate.

1. ProvenanceCan you point to the exact line, written by or about this contact, that proves it?
2. Who wrote itPosts a contact reacted to or commented on were written by other people. They show what the contact cares about. They are never something the contact did, built or said.
3. No stitchingNever join two facts into one claim unless they sit in the same source. A number from one post plus a product from another is a story you made up.
4. FreshnessPosts, metrics and "we're hiring" go stale in weeks. A founding story lasts longer. If the only anchor is old, say so ("a while back you wrote") or drop it.
5. VerbatimNumbers and proper nouns get copied exactly. Never rebuilt from memory.

Ship or cut?

Two prompts to copy.

Paste these into the prompt field of an AI variable (in Vetta: AI Message Settings, in the sequence settings). Change what's in brackets.

Write exactly one sentence, 12 to 18 words, that names the single signal that surfaced this lead, using only the Signals input and the lead's own posts or headline.

Rules:
- Only use facts written by the lead or by their own company. Posts the lead reacted to or commented on were written by other people. Never present them as the lead's own words or actions.
- Never combine facts from different sources into one claim.
- Copy numbers, product names and event names exactly as written.
- If the signal is older than about 60 days, or you are not certain it is about this lead, output nothing at all.
- No compliment, no question, no pitch, no exclamation mark, no em dash.
- Do not mention AI or research. Write the way a peer would say it out loud.

Write the lead's current role as 3 to 6 plain words, for example "head of outbound at a fintech". Use only the job title and company description in Basic details. If either is missing, output nothing. No adjectives, no punctuation at the end.

Here is a message I wrote for cold outreach: [paste your message].

Turn it into spintax. For each sentence that can be said another way, give 3 to 4 variants in this format: {variant one|variant two|variant three}.

Rules:
- Every variant must mean exactly the same thing. If it changes the meaning, drop it.
- Keep variables like {{currentPerson.name.firstName}} outside the curly braces of any spintax group.
- Do not touch my offer, my proof or my question. Only vary greetings, connectors and sign-offs.
- No em dashes, no exclamation marks, no filler like "I hope this finds you well".
- Return only the final message, ready to paste.

Tip: switch Hyper-personalisation off when basics are enough. It's cheaper, and the model has fewer places to go wrong.

Two full sequences. Steal them.

Same four blocks every time: what you noticed, why you're the one saying it, an open question, an easy way out. Hit Preview a send to see how one reads.

LinkedIn: connect, then three messages

Email: three steps

Before you hit publish.

Your ticks are saved in this browser.

• I read the script out loud and it sounds like me.
• Every spintax variant means the same thing.
• Every variable has a fallback or a field that's always filled.
• The AI variable is one sentence with a word cap in the prompt.
• The prompt allows an empty answer when facts are thin or old.
• I previewed at least 10 real contacts and checked each fact against the profile.
• No fact comes from posts the contact only reacted to or commented on.
• The message still reads well with the AI line deleted.
• No em dashes, no exclamation marks, no "I hope this finds you well".
• The first message has no link and no meeting ask.

Edge cases. The part a chat window misses.

A chatbot gives you a great message for one contact. These are the situations where a message goes wrong at scale, and what handles each.

Measure it. Then steal these moves.

• Replies, not opens. Email scanners load tracking pixels before a human sees the message, so opens run high.
• Replies per 100 contacts and positive replies per 100. Those two numbers tell you if the script works.
• Test one change at a time. Duplicate the sequence, change one thing, split similar contacts, compare after 5 to 7 days.
• Read the angry replies. They show you which fact the AI got wrong. Tighten the prompt, run again.
• Write the script first. Add the AI line last, and only if the message needs it.