That email the client roasted
Last week I sent a client an AI-drafted quote. They replied instantly: "Too robotic — doesn't sound like you." I stared at the screen in shock — turns out my old way of using AI for writing was completely wrong.
New approach: counter-examples beat requirements
Later I figured it out. Claude and similar advanced models have a counterintuitive use: telling it "don't write it like this" works 3x better than telling it "write it like this." My friend Ah-Lin runs an independent consulting practice in Hangzhou. Before writing a client proposal, she tosses a "draft a client once yelled at me about" into the AI, says "absolutely don't write it like this," then asks for a new version. Her hit rate went from 40% to 85%. She told me: "I kept pushing AI to write better. Turns out I had to admit where my writing sucked first."
I only just figured this out myself. I used to think AI just couldn't understand plain language — actually, I was the one skipping the counter-example step.
You can try this today
- Money: Claude Pro $20/mo (~$140 RMB); free tier if you'd rather not pay yet
- Time: 1 hour to figure it out the first time, then 10 min per email
- Tech barrier: if you can send a WeChat message, you're set — no coding needed (claude.ai is just a web chat box)
- First step: open claude.ai, sign up, paste your worst-written email in, type "don't write it like this" + the direction you want, hit enter
By stage
Just starting: pick 1 email you edited the most — don't try rewriting all your client comms at once.
Have 1-2 clients: schedule a monthly "AI review day" — paste in client complaints from that month, let it learn your landmines.
Scaling: turn the counter-example library into a shared doc (Notion or Feishu both work) — feed it to AI before each new draft, save those 2 hours of back-and-forth edits.
This isn't for everyone. If your client comms have always been smooth, no rush to try.