Nicholas C. Zakas

Is all of your agent usage metered? Curious how you’re quantifying cheaper.

Yeah, there’s an issue with the terminal that they’re looking at fixing. Not model-specific.

Gave GPT-5 its first real debugging test: pasted logs, it traced the error to its source in the dependency tree. GPT-4.1 chokes on this type of thing.

First day using Copilot with GPT-5 and it's pretty impressive. Only issue for me is that it got stuck in a loop of running the same tests over and over again. The changes were already correct and all tests were passing, not sure what it was doing.

You should! I think more people need to be talking about this approach.

Most of the work on Bredbox (my Pocket clone) has centered around proper metadata extraction from added URLs. After a couple of weeks of back-and-forth with Gemini 2.5 Pro, we finally landed on a solution I'm satisfied with. Now the fun part: getting the UI working.

I'm experimenting with adding a new step to my persona-based AI software development flow: the reviewer. For more complex changes, I'm using Gemini 2.5 Pro to evaluate the implementation against the tech spec to find inconsistencies. So far I'm liking this.

Gotcha. Bug fixes usually don’t require any docs unless there’s a question about the correct behavior. At that point, I’m can be helpful to refer back to the spec defining the feature.

If there’s a departure from the specified behavior, then a decision doc is helpful to leave a paper trail.

I’m not familiar with the term “problem statements” in this context so not sure what you’re asking.

GPT-4.1 has started removing blank lines and comments from parts of my code unrelated to its edits. When I ask it to fix it, it is unaware it has made such edits. 🤔

Thanks. 🙏

Lately, my incoming coaching client applications indicate they learned of me through ChatGPT. It used to almost always be through Google.

GPT seems particularly bad at this for some reason.

Not really the point. The point is whether or not the model obeys instructions the first time or not.

I'd like to propose a new AI benchmark: The number of times you need to say "do not use emdashes" before the model produces content without emdashes.

I'm looking at you, GPT-4o.

The thing I enjoy the most about debating designs with Gemini 2.5 Pro is how it asks me to consider different scenarios when explaining why I'm wrong. That helps me understand its logic and see if my approach works in those scenarios.

Happy "use up all your remaining premium AI requests" day for July!

AI is great, but sometimes a well-crafted regular expression find-and-replace is a better choice.

Thanks to everyone who helped us revise this advisory.

We on the @eslint.org team are also reviewing our policies around issuing security advisories so we can better communicate real security issues.

I ended up writing up a database design spec for Bredbox (my Pocket clone) because I couldn't get one I liked with AI. I then asked Gemini 2.5 Pro to critique it. We debated back and forth on a couple of points, and it was right about everything. Amazing.

If you want to be successful as a software engineer going forward, you're going to need to write and read specs.

Yeah, we worked to get a “low” severity with GitHub. That was pretty easy.

GPT-4.1 threw me a new curveball today: "Because this part requires human judgment, I'll tell you the next steps for you to complete." 🤯

I wish when someone commented on an open source issue saying, "we need this at our company," it also included, "and we're willing to sponsor the project to make it happen." There's still way too much expectation of free labor in open source from for-profit companies.

I don’t have the time or energy to fight, I just want there to be some acknowledgement that dev dependencies running locally don’t represent a significant security issue.

Experimented a bit with GPT-4.1 Beast Mode and ultimately found myself falling back to my PRD -> tech spec -> implement approach. 4.1 just couldn't figure out the complete picture for some fairly basic features.

Do any of my followers work at Snyk? I'd like to talk to someone about this report, as I think it's unnecessarily alarmist given how this package is actually used:
https://security.snyk.io/vuln/SNYK-JS-ESLINTPLUGINKIT-10847878

Yeah, actually tried a new footer but didn’t like the result. Will try again next week.

Before: Screenshot of the Human Who Codes website homepage showing the main navigation menu (Books, Reading, Coaching, Newsletter, Donate, Contact) and a Recent Posts section with blog post titles and excerpts including 'A persona-based approach to AI-assisted software development', 'Post to social media using Claude Desktop and Crosspost', 'Introducing Mentoss: The fetch mocker', and others. The right sidebar displays book covers for JavaScript titles and a Recent Snippets section with links to various technical tutorials.

After: Updated view of the Human Who Codes website showing the same layout but now displaying individual blog post cards in a more detailed format. Each post card shows the full title, description, author photo of Nicholas C. Zakas, publication date, estimated reading time, and topic tags. The first post about AI-assisted software development shows tags for 'AI', 'Claude', and 'GPT'. The second post about Crosspost shows tags for 'Open Source', 'AI', and 'Claude'. The third post about Mentoss shows tags for 'JavaScript', 'Open Source', and 'Fetch'. The sidebar content remains the same with book covers and recent snippets

I unleashed Claude on my blog and told it to "make it prettier." Here's the result. (now live at humanwhocodes.com).

Copilot agent spent a half hour investigating this issue and figured out it could make a one-line change to solve the problem. That's exactly what a good engineer would do. https://github.com/humanwhocodes/tailwind-csstree/pull/14