I have a weekly task that runs in CI and finds all our flaky specs that have been failing randomly over the last week. An AI agent analyzes the report and fixes them all. This is at least weeks of painful debugging work if I had to do all this myself. And it's all done for me in the background while I work on other things.

I just asked Astra to go through all our blog posts, fix or replace all our banner images so that they're all the same size, review all the posts to fix any mistakes or add missing info, and make sure they're all using the right categories and tags. I love how you can give it access to the Ghost API and it will just take care of everything. docspring.com/blog/ looks a bit better now

LogStruct is now one of most well-tested and bug-free libraries I've ever built: github.com/DocSpring/logstru… Over the last week I've let agents work on it semi-autonomously in a loop. I gave them the instruction to just keep going until there's nothing left to fix. It initially took 10 rounds of review from Opus 5.5, and then 9 further rounds of review alternating between Opus 5.5 and GPT-6 Astra. It also included 9 rounds of staging QA while testing the library with our own code. Plus the fact that I've been digging into our production logs over the last few weeks, which is where all of the original issues were found. The library is also fully typed with Sorbet, and the docs are tightly integrated with the code. I even type-check all the code examples that we use in the docs. So I'm pretty confident, but I won't claim that it's perfect just yet. It still needs more people to use it in their own app and send feedback, and we might find some more issues over the coming weeks and months while we monitor our production logs. I also wouldn't be surprised if there are still a few mistakes in the docs.

I thought it might be annoying, but the automatic copy-to-clipboard feature in Herdr is really nice. Just select or double click some text and it goes straight into the clipboard

8 rounds of PR review with 4-5 subagents each on Opus 5.5, including full QA runs on staging checking the browser and logs. GPT-6 Astra still found 11 issues that Opus had missed: 3 high, 3 medium, 5 low

current status

Probably going to turn off Coderabbit. It was useful for a little while but it can’t compete with our own code review skills now. I often run 5-10 rounds of code review and alternate between Opus 5.5 and GPT-6 Astra. Coderabbit finds almost nothing compared to these reviews

Enterprise customer asked for a quote. I sent them a quote. They sent me a purchase order. I sent them an invoice. They sent me a check in the mail to our US postal address. I paid to courier the check to NZ since I can only deposit through the Mercury app if I have the physical check. We’re living in the future, but I think the future will still have physical checks and fax machines for a little while.

Opus 5.5 is great but it definitely pays to double check everything with GPT-6 Astra. Here's an example

I've got a bunch of Claudes in my browser

I don't actually think I want to automate this or do it in OpenClaw/Hermes or some app. I think Claude Code is the best UI for this, and I really like to be in the loop the whole time.