by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
AI output quality depends on who absorbs the rework.
More than 300k people saw a diagram from my friend Clare on my feed last week about AI. A few of them told me it didn't apply to them. The one that stuck was Jennifer.
Jennifer is a project manager; her point was that my post assumed that she has time to think, but that's rarely the case. She uses AI to gather data quickly, and since she always has tight deadlines, it works well.
She was right about the impact of deadlines, and it took me a day to see she was right about something bigger.
Clare's diagram has three bars, and most people's mental model focuses on the first two. Without AI and with good use of AI. Both take more or less the same time, but with AI you achieve a better outcome. Nobody talks about the third and longest: start in a rush, AI produces, rework, verify, and then someone cleans it.
And this is the point that I keep seeing: the cost of AI "speed" rarely falls on whoever caused it. It lands on someone down the line a few days after, and that rework is almost never attributed back to the source.
Workday and Hanover measured it and published it in January this year. They surveyed 3,200 active AI users at companies with revenue exceeding $100 million. In that survey, they found that from the 10 hours gained due to AI, almost 4 hours are wasted correcting the work.
Now let's get back to Jennifer; she thought she was telling me the diagram didn't apply to her. She wrote the reconciliation prompt for the person who runs the process: her.
That's what the diagram doesn't show you. Jennifer has no rework because she's not delivering to anyone else. She eats her own output, and AI quality depends on who absorbs the rework. If the producer and cleaner are the same person, the incentive to deliver quality is very high.
Most times people are producing for someone else. That's why the speed feels free: "someone will check it" is true, and that someone is not you.
If you're building for someone else, consider that delivering as fast as possible is not the best approach for high-quality final results. Someone will be very unhappy fixing what you did.
The Build Log
Google has been telling me that thoughtled.ai was fine, but for all other answer engines it has been blank since May.
I found it completely by chance; I was testing the new Deepseek Harness (dsh). Try it yourself:
1. run npx @deepseek-ai/dsh web in the console
2. Open http://127.0.0.1:3080/ in the browser.
I asked it to help me improve copy on one of the pages, and it told me there was none. Our website was built in Lovable months ago and uses client-side rendering. Since Googlebot can execute JavaScript, everything looked fine in Google Search Console. GPTBot, ClaudeBot, and PerplexityBot only read the HTML, so they have been ignoring it for three months.
An extremely simple problem that, of course, Claude could have also found, but I had never asked it to look.
Dsh didn't win on capability, it won because it was already running. DeepSeek launched the harness with an MIT license on 13th of August, and today it has 184 thousand stars.
In dsh, everything is a plugin; you can replace anything with your own version, and the best part is that you can see all the stack traces, how the tools were called, and how and why the decisions were made.
You can even use any model you want. I've used DeepSeek V4Pro and GPT-5.6.
It took 2 commands and 10 minutes to run an initial dsh test that helped me find a bug that had been there for 3 months. The fix took 2 minutes: I just asked dsh to implement prerender.
If you're building, try dsh and make sure your marketing sites are readable by search engines. And if you're not building consider give part of your AI work to DeepSeek, you'll probably be surprised.
On My Radar
The August outage was caused by a bad deploy. Monthly commits have doubled from 1.4B to 2.9B since April. Agentic coding is here.
Agents hit 94% on Terminal Bench 2.1, but the result was fake. If your agent is performing too well, make sure you audit the results.
PointFive ran the same prompt in 2,908 sessions with Claude Code, and in some cases token reduction increased costs by almost 50%.
Enforce an AI agent's permissions throughout the entire task, not just at the start. The runtime stopped or corrected 94.8% of rule-breaking actions.
Adoption across tens of thousands of teams. The number about what people actually do, not what sales departments want you to think
Need another pair of eyes on your AI project?
One consulting slot available for August → Book a call
Thanks for reading.
See you next week. Go build something.
Luis
Know someone who'd enjoy this? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
