Does anyone have this issue crop up from time to time, especially with Claude Code? I know that it's basically impossible to rely on the model to "magically know" or infer what you truly mean. And that structured prompting is important. But does anyone else ever find that Claude Code, and sometimes Codex goes from "Can give it sensical instructions and clear plans and they work" to "Forgets clearly stated plan details mid-run. Orchestrates dynamic workflows in the most retarded way possible. Generally just becomes unhelpful and makes you suddenly realise how powerless you are if they just decided to increase a Claude Max sub to £5K per month" my harness/agentic engineering skills are not god tier, but there's a noticeable variation in the quality of what i manage to get it to do some days, where the output "matches my brain" vs other days where it feels like they might have trained the model on every single psychological weakness i have to make me do enough dumb shit to warrant spending more money on fixing stuff that was previously working not sure what to do about it either - seems like it's a common issue with Claude Code/Claude in general i feel like it's just a sign that my harness needs to be more robust, so the model cannot fuck stuff up as easily. but idk. @Corbin Brown do you deal with this at all and if so how do you un-fuck it?

Hey guys - @Corbin Brown I've seen you mention the idea of QA/Staging environments a few times. Going through this process for the first time with an app I hope to deploy to production soon. I am curious how you structure your QA/Staging tests? Currently I have a staging branch on Vercel, and I've used Codex High to look at all the app's features and create a matrix that specifies what to test for each feature/component of the app, how that test is to be performed (Playwright, Cursor Browser etc) and what success criteria are. Then Composer has gone through and performed all these automated tests. Some are still manual that I need to do. Composer's basically found a lot of bugs/errors, which it's fixing in the staging codebase. My plan is then to do this, finish the manual tests and then deploy to production. Is this general an acceptable way to run QA/Staging or am I missing something mega here?