Tuning AI messages with skill files

The prompt of an AI variable says what to write. The skill file says how you sound and what is true about your business. When a generated message misses, the fix is often one line in the skill file rather than a longer prompt, and every variable that uses the file improves at once.

Find the skill file a variable uses

Open the variable in AI Message Settings (Configure variable) and expand Skill files. The attached documents are listed there; Add business context attaches more. See Setting up AI variables.

Edit the skill file

Skill files are business context documents, so they are edited in one place:

1. Go to Settings and open AI Business Context.
2. Click the document in the list.
3. Edit it in the markdown editor. Headings and lists help the AI scan the file quickly.
4. Click Save.

The change applies to every AI variable, agent and AI feature that uses the file, the next time it writes. For the full picture of business context, see Business context and AI Skills.

Because one file can feed many sequences, an edit changes all of them. When a change is meant for one sequence or one agent only, create a second document for it and attach that one instead.

When a change takes effect

A sequence writes each lead's AI messages just before their first send, and then keeps them. So a skill file change applies to every lead whose messages have not been written yet, which in a running agent is usually most of the queue.

Messages already written for a lead, including the ones you previewed, are not rewritten on their own. For those leads, open their messages in the preview and regenerate them, or edit the steps that have not been sent yet. Messages that have already gone out cannot change.

What to put in a skill file

Three things make the biggest difference to generated messages:

• Voice guidance: how you write. Sentence length, tone, words you use, and a short never list ("I hope this finds you well", exclamation marks, "quick question").
• Proofs: short, specific results you can stand behind. "BrightLedger: close went from 9 days to 3 in the first quarter." Numbers beat adjectives.
• Case studies: one or two paragraphs per customer: who they are, the problem, what changed. The AI picks the one closest to the lead.

A simple structure:

# How we sound

**VOICE:** plain, warm, direct. Short sentences. One idea per line.
**NEVER:** "I hope this finds you well", "quick question", exclamation marks.

## PROOFS
- BrightLedger: close went from 9 days to 3 in the first quarter.
- LumenPay: one workspace for spend, close and cash across 4 entities.

## CASE STUDIES
### BrightLedger (fintech, 200 people)
...

Draft it with ChatGPT or Claude

If you already have the material (a sales deck, call notes, a few emails that got replies), paste it into ChatGPT or Claude and ask it to summarise it into the structure above: voice rules, a never list, five proof points and two short case studies. Then edit the result, paste it into the document and save. Keep it factual: the AI quotes and applies what is written, so vague copy in means vague messages out.

Then check the result

After saving, preview a lead and regenerate its messages to see the new copy: Previewing and regenerating messages. Try two or three different leads, since a rule that fixes one message can read oddly in another.

Common questions

Key outcome

A tight skill file gives every AI variable the same voice and the same facts. Spend your tuning time there first, and keep prompts short and specific to the message they write.