Published: 2026-07-18
I've been building a lot with Codex recently, and it's been great. The new GPT-5.6 Sol is a really powerful model, it does so much on it's own. To the point where it almost does too much. It wants to validate everything every time. Which is great, but incredibly costly. It writes tests, it runs through the web site, it deploys. One thing in there is repetitive and costly. Running through the website.
To me a project is just a list of ideas: the site allows conversations between seller and customer, you can sign up and get a welcome email, you can upload images and then send a message to a customer and the image shows up. Now combine all of those. And yeah, these could be done in a playwright test, but then you're managing a playwright test. I don't really want to do that, projects are flexible, they're flaky, they can't adapt to small changes.
I've started writing markdown files called "ideas". They're really basic, describe an idea for your project, give the idea to your favorite coding agent, have the agent verify the idea is done, if it's not, tell it to make sure it gets done. It works surprising well. I've gotten a lot of value out of just these simple files and skill. GPT-5.6 Sol loves being given a goal, and it will go through the entire thing to make sure it's true. Given the right harness and freedom, it will do everything.
I've been seeing a lot of blog posts about what it means to build things, and it's adapting. I no longer really look at code... well I do and don't. I really care more about how it works (I've sort of always cared more about that), than how it's implemented. If the feature works, if the idea is verified, why should I really care about whether the underlying code is the best it can be. I've been trusting these agents to have my back more and as they improve, their output is getting better.
I can write an idea that says "The Metner project should have a strong, welcoming design that is attractive to micro-preneurs." and weekly it can go through and make sure that stays true. While I have my day-to-day updates, this will keep things in line, or give me tasks to update to make sure it stays in line. They're also living documents, they can change as my priorities change. If I no longer think that an AI message that's drafted needs human approval, I can modify the idea to make that happen.
Anyways, if you want to check out the skill it's located on my Github. Pull requests are open, feel free to contribute, or open issues. I'm not fully set on this, but I've seen some blog posts saying "humans own the ideas, agents own the implementation", and I wanted to make something that actually follows that.