by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
Alex Cui sells AI detection software for a living. This week he wrote on X why the thing Anthropic just shipped probably can't do what people think it does.
The internet went crazy this week when people realized that Anthropic started adding a watermark to Claude's output at the start of August.
Internet being the internet, 1 day after the announcement, a GitHub repo claiming to remove AI watermarks was already going viral.
As crazy as it sounds, we're in an AI watermarking arms race.
The problem is that it's clear what the final chapter will be. At ICML 2024, there was a paper that showed it's impossible to do. Anyone who can distinguish good writing from bad can reword it and strip the watermark.
In his post, Alex Cui discusses a use case that should worry builders more. Code, unlike free text, has strict rules of what goes where, and you can't just swap words and keep the meaning.
My question is whether the coding models will keep the same quality, or whether, because of this watermark, the quality will reduce.
Maybe Anthropic will just change the variable names or the comments. If that's the case, the GitHub repo already includes a script to rewrite the comments and rename variables while keeping the same names in the public API.
If you're building, the mark isn't the reputational problem since everyone assumes you're using an Agent; hopefully the quality will be the same.
If you're not building and you're passing Claude's work as your own, you're running out of time. To avoid the watermark, you'll need to actually do some work and rewrite the content.
Together with Krisp.ai
It sits in front of your speech-to-text, isolating the primary speaker from competing voices.
Across their test set, word error rate drops by 46% on average. In the hard case of a second person talking, it goes from 35.9% to 10.9% across 10 STT systems.
The Build Log
I rewrote the campaign modules of thoughtled.ai last week: all the content creation is now assisted by a content strategist agent.
For this work, I completely changed my flow. I used Claude to make the changes and run the tests. Then handed it over to GPT-5.6 Sol for review and fixes, and finally I did the final review myself.
GPT-5.6 Sol found 12 bugs in corner cases that Claude didn't see, and I probably would have missed some too.
I've been running Claude Code in auto mode for a few weeks (after agreeing to a plan with it). On 14th of August, Anthropic made auto the default. Their argument: manual approval stopped being a real review a long time ago because users approve 97% of permission prompts.
When you remove step-by-step approval, you need proper guardrails. I have hooks and a very complete e2e testing framework, but GPT-5.6 was still able to find some bugs.
That's why using a completely different agent helps. Different models, different harnesses, and also different assumptions about what fails.
So my setup now is two agents: Claude for the initial development, and Sol to check the diff and report back if there are any problems to solve. Both are in auto mode and remain engaged until Sol is happy that Claude did a good job.
The quality increases, but so does the cost in tokens. The big advantage is that I can do other stuff when a feature is being written, and I just need to look at it for the last code review.
I'm back to having a team producing features, and I'm the senior engineer checking their solution.
If you're not building yet, take the last work you did with one model and ask another to cross-check it. I've been doing it for a long time, and the quality of the final documents has become consistently better.
On My Radar
Proprietary reasoning can be recovered from its encrypted traces. Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models.
Claude's watermark probably doesn't work how you think and might impact your code quality.
Routing between NVIDIA Nemotron 3.5 Lightning and Claude Opus 4.8 cut the total cost by 74% against running Opus alone, while retaining 93% of its accuracy.
The hard part is managing, evaluating, and improving the agents. Verification is the real constraint on automation.
Agents might be the primary users of software soon by sheer volume. For you the important screens will be "approve", "undo", "show me what changed"
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/
