💎 While many are building super-systems with multiple skills and sub-agents and then running Claude Code for hours, I'm going the opposite direction — even more meticulous.

Yes, Codex-5.3 and Opus-4.6 are good, but still not good enough (in my opinion. But you'll understand why below).

What is my framework?

Suppose we start a project from scratch or are about to build a super-large feature.

1. I discuss the implementation with GPT and Claude in parallel for a long time, periodically sharing ideas between them. Then I stick with Claude to have him create specs, and I show each spec to GPT. He finds a lot of imperfections! And together, the three of us refine the specs.

2. Then, based on the specs, we create a plan file where each section is detailed as much as possible by tasks, and it specifies which specs the AI agent should refer to. Also, it's necessary to update CLAUDE.md and AGENTS.md for the feature/project. They must specify that tests and documentation are required for each module!

3. Then, for convenience, Claude creates prompts for me that I just copy-paste and hit start at each stage.

4. Then, in separate chats with Claude, we start working on the specs. 1 chat = max 1 section. Once the base is ready, we can parallelize. For each section, the Claude coding agent first makes a plan, I show it to Codex, he validates it and usually points out many weak spots. Usually one iteration of plan refinement is enough, but sometimes 2–3. Once Claude finishes the implementation and gives a summary of what was done, I show it to Codex again, he finds weaknesses and bugs. Then Claude refines them. Then /compact and continue (or a new chat).

4.5 Yes, I use Codex here for a reason. If I do a review of the Claude agent in a separate chat, it works too, but worse. And I use codex-5.3-extra-high here.

5. After each section, and even better — more often, I need to do another review via the /review agent. Here I use both sonnet-4.6 and Codex-5.3-high (without extra), and they find different weaknesses. You can also run a security-review in Claude if the features might pose security risks.

6. And so on until the specs are completed. Then you can optionally run the agents through them again to check if everything is done correctly.

Additional points:

— Attention: I at least skim through what the agents planned! Otherwise it's a gamble: the longer you do this, the higher the chance of significant deviation from what you have in mind. Even with people, constant synchronization is needed, let alone with agents.

— It's also advisable to manually test and see what was actually done, not just rely on reports and test passes.

— The framework above may be overkill for an MVP that needs to be thrown together in 2–3 hours. But if you have time and tokens, it's better to do it right from the start, because refactoring later can be longer and more painful.

P.S. I use 1code.dev
They have both Codex and Claude Code. They've already polished the solution quite well; I bug-reported as much as I could because it's really convenient. More convenient than the naive Claude Code desktop. The only thing is, be prepared to have 16GB GPU if you want to run 4+ chats simultaneously.
Also, it's nice that the founder is also one of us, from Russia.

What frameworks or lifehacks do you have?

The video above is just to grab attention) It's relevant to Anthropic's research on which professions will soon disappear)

#vibecoding

@oh_my_zen