Drafting the first reply to each new sales lead

Who it’s for

She works in sales in the travel and hospitality industry. When her company assigns her a new lead, she sends a first email that asks the questions she needs answered before she can suggest a trip: how many people are going, when, roughly what budget, and what they enjoy on vacation. She has always written these herself, one at a time.

How it works now

First thing in the morning, she opens Claude and gives it the names of the leads she wants to reply to. With each name she adds any notes about the request that Claude wouldn’t otherwise have. For example, her company doesn’t book Airbnbs, so if someone asked about one, she says so.

For each lead, Claude reads the lead details in her Gmail and drafts a reply. It asks only the questions the lead hasn’t already answered, so someone who said they’re a party of four isn’t asked how many are traveling. It leaves each reply in her drafts. She reads them, edits what she wants, and sends them. Claude’s Gmail connector can create drafts but can’t send email, so nothing goes out without her.

What changed

On October 19, 2026, she replied to 13 new leads this way in about 30 minutes. Thirteen used to take her a little over two hours.

The replies get about the same response rate as the ones she wrote by hand. The difference is the time she gets back.

What we built first

The first version ran on its own every morning as a scheduled task in Claude Cowork. It searched her inbox for new lead notifications, drafted a reply to each, and labeled each notification in Gmail so the same lead wasn’t drafted twice.

She has turned the schedule off. The scheduled version only knew what was in the lead details. It had no way to know things she knows about a request, like the Airbnb example above. Writing a list of names with a few notes takes her a few minutes, and Claude still writes the emails.

The setup

Both versions use Claude’s Gmail connector, and the first used Cowork’s scheduling. Both are part of her existing Claude subscription. It’s a business plan, so her conversations and emails aren’t used to train Claude. Is it safe to put client information into AI? explains why the plan matters. Connecting Gmail and writing the instructions took a few minutes. There’s nothing to host and nothing extra to pay for.

Making it sound like her

Most of the time went into a brand voice skill: written instructions that Claude reads before it drafts, describing how she writes. We spent about four hours on it.

• We collected emails she’d written that got a response, as examples of her voice.
• We added a few that didn’t get a response, to see whether Claude could find a pattern that separated the two. I don’t think that part worked or added much.
• Claude interviewed her about how she writes and what she wants her emails to do.

The skill came out as a short list of rules. Open with something specific to the client’s trip. Make one point or ask one question per email. Say so plainly if availability is limited, but never invent urgency. Close without pressure. It also lists phrases she never uses, such as “I hope this message finds you well” and “just checking in,” and it fixes her sign-off.

Keeping the emails short

I think a large part of why this works is that the emails are short. A few sentences that sound like her are quick to check, so she sends them.

What isn’t handled yet

The notes she adds each morning cover cases the skill doesn’t know about yet. When the same kind of note keeps coming up, it can go into the skill, so she doesn’t have to type it again. We haven’t done that yet.

What it doesn’t do

• It doesn’t send anything. She does.
• It doesn’t choose which leads to reply to. She does.
• It doesn’t follow up when a lead doesn’t answer.