tldr: (next message) Details: PART 1: 1) I have listened to every episode of AITW. The opinion I had formed was we can't just vibe code (as transformer-based coding agents have limitations) - and we need to review the plan/outline/code. 2) Peter Steinberger says chats back and forth with AI, reviews the code and refactors then once once a week or so - this was way back in October or so 3) I am not a senior SWE and have done a lot of vibe coding for complex greenfield apps (MVPs and V1s) using $200 Claude and Codex - due to time constraints, I haven't reviewed the code yet - Note: I understand the project structure and modules, and verify that the plan (from plan mode) is not off by a lot. 4) Claude has gaslighted me so many times that I switched over to Codex - my mental model: both have same limitations of transformer architecture - but Codex is methodical and processes longer so works better. PART 2: 5) In recent Ryan Lopopolo of OpenAI's interview on Latent Space he talks about the Harness engineering where humans wrote 0 code - given he's from OpenAI he'll probably hide some of model's limitations. 6) DHH in a recent podcast on Pragmatic Engineer - says generated code is correct most of the time (that's a lot coming from someone who supposed to be obsessed with code quality) - and both he and he host agree on the step change in coding capabilities since late-November 7) The last 2 points above have shifted my perspective - that the latest coding agents (esp. Codex on high/xhigh) can write decent code - and for founders building V1s (that they plan to extend to more features after some code review/refactor), the whole RPI+ (I forget the new abbreviation for the 5+ step process) process is a overkill. Also, Dex is not giving me access to Riptide to try it out. lol. So I haven't been able to evaluate that process myself. @Vaibhav (BAML 🐑) @dex Thoughts?

Although Peter Steinberger talks about how he refactors, if we consider OpenClaw as a product, I think it falls short; however, as an open-source project, it excels in its design philosophy

great thoughts! here's how i think about it:

Yep. There will always be some “frontier” that is available if you just yolo prompts, and your willingness and skill to context engineering will define your ability to push past the frontier

Thanks for the reply @Vaibhav (BAML 🐑) @dex . Pls do answer this follow up question : Let's say - we specify everything (as much as possible) by going through the RPI+ process and - we are working on something so new that it's not in the training data or - it's diametrically opposite to statistical patterns that it trained on, - from your usage of RPI+ method - do latest coding agents implement them perfectly? NB: An analysis from HumanLayer's and BAML's ticket implementation history, would be great. That's probably asking too much. Unless you guys want to publish a research paper. lol.- btw, it would be freaking awesome and useful.