People say a longer prompt makes better code. It does not. A short prompt plus the error log beats a page of rules. The model copies whatever you put last. What do you put at the end of yours?

Devs use AI to write code. Non devs use it to get the work done. Which one are you, and what did you ship with it this week?

You haven’t truly seen the creator economy at scale until you visit China. 🇨🇳 Every single scenic spot, mountain trail, and lake view here isn't just full of tourists it’s full of tripods, ring lights, and live streamers talking non stop to their phones. At first, you think: "Why is everyone live streaming their vacation?" Then you look closer: It’s not vanity. It’s pure commerce. 1. Real time Live Commerce: They aren’t just sharing the view; they’re selling discounted hotel vouchers, cable car tickets, and local tea right off the screen. "Look at this lake behind me tap the yellow cart to book the resort next to it for 40% off." 2. Cloud Tourism (云游览): Millions of desk bound office workers in tier-1 cities tipping streamers live just to tell them: "Pan left, walk toward that cliff, show us the sunset." 3. The Proof of Presence (打卡) loop: Algorithms here ruthlessly favor hyper local, live verified feeds over edited VODs. The West talks about social commerce as an experiment. In China, nature reserves literally run on bandwidth infrastructure designed for live monetization. Wild to witness in person.

Paste the failing test into the first message, not the feature name. Claude will match the test faster than it will match your description. If the test is green before the change, the task is already wrong. What do you paste first, the error or the file?

Claude Code now hides the lead message to the subagent. That is the wrong default. If I cannot read the handoff, I am reviewing a result with no spec. The trace is the job. The diff is the leftover.

You Should Know is an admission. The main agent output is not the review. A side agent that only scans the text still cannot fail the run. I do not need a recap of the diff. I need the claim with no test, the file it was not allowed to touch, and the assumption it treated as fact. If the mod cannot block the commit, it is a newsletter in the terminal.

Bigger AI diffs are not speed. If the PR grew and the test gate did not, you only moved the work to review. What fails the build before you even open the diff?

“AI writes the code, I just review” is backwards. Review is the job now. If the PR got 3x bigger and you have no test gate, you did not get faster. You delayed the work. Where does review break for you: size, or missing tests?

What’s one boring work task you handed to ChatGPT or Claude this week? Did it actually save you time?

“AI writes the code, I just review” is backwards. Review is the job now. If your PR grew 3x and you have no test gate, you did not get faster. You deferred the work. Where does review break down for you, size or missing tests?

Put one file the agent must read every run: the test that has to pass, and the path it may not touch. If that test is red, the task is not done. What do you pin first?

sharp take for developers A green test that broke login is not a win. Pin the exact test command and the one file the model may not edit. Vague “keep quality up” is how auth dies.

Before Cursor or Claude Code touches a branch, paste two lines: the exact test command that must pass, and the one file it may not edit. That stops a “green” patch that breaks login. What is the first command you pin?

Those are the AI image takes to my Jiuzhaigu china trip. Let's see what you think is better?

simple tip for non-developers Do not ask ChatGPT to “write the email.” Paste the messy notes, the names, and the one thing you need the other person to do. Then cut the first and last paragraph. What work task do you still type from scratch?

Not much AI testing I've been able to do these days but had to share some of the stuff AI image generation just can't make yet. Let's see if any of you can make these type of images just with AI? @grok can you create similar images? All images are real as part of my trip to china

Before you let Claude or Cursor touch a PR, pin the acceptance test in the prompt: one command that must pass, one file it may not edit. Vague “fix the bug” is how you get a green test and a broken login. What is the one command you always paste first?

What’s one boring work task you handed to ChatGPT or Claude this week that actually saved you time?

AI is already changing China. When looking in the streets and talking to people I can see they all talk about DeepSeek for their work. You can see also in the pricing that the peak sessions are Chinese times and the whole Chinese holiday is off peak. They use it every where, from talking to foreigners to analyzing images in stores and cameras around the streets. Will be interesting to see it in 2-3 years as the evolution is just getting stronger! Any other country use it that much already as part of their day to day?

AI is already changing China. When looking in the streets and talking to people I can see they all talk about DeepSeek for their work. You can see also in the pricing that the peak sessions are Chinese times and the whole Chinese holiday is off peak. They use it every where, from talking to foreigners to analyzing images in stores and cameras around the streets. Will be interesting to see it in 2-3 years as the evolution is just getting stronger! Any other country use it that much already as part of their day to day?