Claude Reply Drafts: Review Before You Hit Send
Many owners start the morning like this: open the phone, find a dozen customer messages on LINE and three price inquiries in the inbox, with the Excel price list on another computer. Before the coffee is finished, an hour has disappeared.
For this week's Claude mini project, I want to help you build a "morning customer inquiry triage + Claude reply drafts": AI sorts the inquiries and drafts replies first, and you just review, edit and hit send. The inspiration comes from two articles Anthropic published over the past two weeks, and an important Cowork update this month.
Today's Highlights: Claude reply drafts. One sales rep's six AI agents
The case: a sales rep in Japan hands two morning hours to AI
On October 16, Anthropic shared how a business development representative on its Japan team works. He used to spend one or two hours every morning assigning and screening leads and writing emails in Japanese or English, and the CRM was often updated a day or more late.
He then built six AI agents, each responsible for one thing: one assigns new leads to the right colleagues every day; one grades his own leads, updates Salesforce and drafts the first outreach email; one checks the inbox every hour and drafts replies; one writes customer interactions back to opportunity records; one summarizes the day's meetings and drafts follow-up emails before he leaves work; and the last adjusts the other agents' instructions at his request.
A few details worth noting: every draft comes with a short background summary, and he reviews and edits each one himself before sending; pricing questions, sensitive customers and ambiguous requests still go back to a person. Based on his weekly work logs, he estimates the agents save him about 60 hours of manual work; but he also admits that getting the drafts to sound like him took hundreds of rounds of "draft, correct, save." He also found that a shared knowledge base affects draft quality more than how long the prompt is.
A reminder from the engineering team: the worst thing a scheduled task can do is pretend all is well
Another article on scheduled automation, from October 8, lays out several practical principles. First, pick up each run from where the last one left off, rather than always grabbing the past 24 hours, so nothing is missed or duplicated. Second, when a source fails to load, say so clearly instead of treating it as a quiet day. Third, double-check every item before sending; one outdated alert hurts trust more than missing a few. Fourth, set a spending cap per run, and give read-only access wherever only reading is needed.
On the tools side: Cowork scheduled tasks now run in the cloud
The Claude Help Center notes that since October 6, new Cowork tasks on Pro and Max plans run in the cloud, so scheduled tasks keep running even when your computer is off, and progress carries over between web, desktop and mobile. However, reading folders on your computer still requires the desktop app to be open, and is limited to the folders you've connected.
Jesse's Take: Let AI draft first. Don't rush to let it send for you
An honest word first: this case comes from Anthropic's own employee, whose job is selling Claude, and the 60 hours is his own estimate. The cases vendors choose to share are usually the best ones, so borrow the approach, but don't hold yourself to the numbers.
Still, its structure suits small businesses in Taiwan well. A common scenario: a bakery ingredients wholesaler whose customers ask for prices on LINE, place orders by email, and occasionally call to complain about delivery times, with data scattered across LINE, Excel and the owner's head. Here's how I'd suggest getting started:
Suggestion 1: Do just one thing. Sort first
Don't build six agents from day one. Ask Claude every morning to sort the previous day's inquiries into three categories: price inquiries, after-sales and complaints, and other. Sorting alone tells you who to reply to first today.
Suggestion 2: Write a one-page "reply knowledge base"
Put price ranges, delivery times, shipping rules and common questions on one page and add it to the project. The case shows this one page is more useful than agonizing over your prompt.
Suggestion 3: Quotes and complaints always go back to a person
Drafts can be handed to AI, but keep the "send" button in your own hands, especially for emails involving money or emotions. Every time you edit a draft, add "why I changed it this way" to the knowledge base, and AI will sound more and more like you. LINE messages currently need to be exported or pasted in by hand, so don't let that step put you off.
A first step you can take right away: Before you leave work today, pick 20 customer inquiries from the past two weeks, paste them into Claude, and ask it to sort them and draft three replies. Edit each one, then write your editing principles on one page. Next week, turn it into a scheduled task that runs every morning.
Further Reading
Related posts on JCXO
Original sources
Comments