@pro.glitch: Build an AI employee that answers your customer emails — in Claude Code, every step What it looks like working: overnight, 11 customer emails arrive. By the time you sit down, 9 have replies already written in your words, with the right prices and the right policy quoted. The other 2 sit at the top untouched and flagged — one person is angry, one wants money back, and those are yours. You read, you approve, you're done before your coffee cools. THE BUILD: 1 — THE DESK. Make a folder, open it in Claude Code, and write the job description in one line in CLAUDE.md: "you draft replies to customer emails for my review." One employee, one job. 2 — WHAT IT NEEDS TO KNOW. Plain files: what you sell and what it costs, your policies (shipping times, return window, what's covered), and the questions customers actually ask — which you know by heart. Help page or FAQ goes in too. This is what it pulls real answers from instead of inventing them. 3 — YOUR VOICE. The step people skip, then wonder why drafts sound like a robot in a call centre. Go to your sent folder, pull 20 replies you were happy with, paste them in a file. Twenty real examples teach your rhythm better than any instruction you could write. 4 — CONNECT THE INBOX, CAREFULLY. Grant READ. Grant DRAFT. Do NOT grant send. Delete stays off. A draft you don't like costs three seconds; a sent email you don't like costs a customer. 5 — WRITE THE PROCEDURE → SAVE IT AS A SKILL. Read the email, work out what it's about, find the answer in my files, write it in my voice, quote the real price or policy, and say which file it came from. Plus the instruction that matters most: if you can't find the answer in my files, SAY SO in the draft instead of guessing. Then: "save that procedure as a skill called customer-replies." 6 — THE ROUTING RULES (what makes it safe to leave running). Anyone who sounds angry → straight to you, no draft. Refunds and chargebacks → you. Anything legal → you. Anything it's unsure about → you. Everything else gets a draft. You're teaching it the difference between a question and a judgment call. 7 — TEST IT ON YOUR OWN HISTORY. Take 20 emails you already answered last month, run it on those, compare its draft to what you actually sent. You'll see exactly where it's wrong — an old price, a customer who'd already emailed twice, too formal. Every fix goes into the SKILL FILE, not into your head. Repeat until the drafts stop surprising you. 8 — LET IT RUN ON REAL MAIL. It drafts, you review. Your job becomes reading and approving instead of writing from scratch. Every correction from here follows the same rule: if it'll come up again, it goes in the skill. THEN, OVER WEEKS: as you watch it be right, you widen what it's allowed to do. The simplest category — the one it's been right about 40 times running — is the one you let send by itself. Money and angry customers stay gated behind your approval forever, because trust earned on easy questions doesn't transfer to hard ones. That's the build: one folder, its knowledge, your voice, a careful connection, a written procedure, hard routing rules, and a test against your own history. #aiemployee #aiautomation #aiagent

Professor Glitch

Open In TikTok:

Region: US

Friday 18 September 2026 20:00:00 GMT