Using ChatGPT for Planning vs Codex?

Codex
1

Good day.

First, I’d like to say that Codex has become such an incredible, powerful tool. I hardly write code anymore - what a paradigm shift! It’s remarkable to look back and recall how seamlessly we have transitioned.

Naturally, cost and power intertwine. We spend moar, and/or strategize to reduce cost.

As a result, I have been using ChatGPT w/ GitHub connector to plan and debug, while using Codex to execute, and have been wondering if anyone else has done so?

Advantages

  • Doesn’t cost Codex credits
  • Is much easier for back-and-forth
  • Can be used in any typical conversation

Disadvantages

  • I imagine the model isn’t as suitable as Codex, and more likely to make bad decisions? However, I haven’t really been able to notice any difference

Wonder if anyone else has adopted this strategy, and what they’ve seen.

I imagine this would be effective for many use-cases - not strictly coding.

2

This has been my go-to approach for complex tasks for quite some time.

I use a Pro model for planning, then Codex for execution and implementation. I also often circle back to the Pro model for review.

It is a time-intensive process, and I have to admit I have not tried 5.5 Thinking Heavy extensively yet.

The Pro models are just too good.

3

Exactly! Its ability to review the branch and check alignment has worked out really well. I’ve always found the Codex planning mode (in fact, all planning modes) to feel…awkward.

Arriving to the party late is fashionable anyways… Right? :joy:

4

Which gets me the idea to use planning mode to explore topics instead of simply chatting or actually thinking about it.

Create a plan to explore this idea. What should be investigated first …

It’s never too late to exchange learnings!

5

I’ve tried with ChatGPT written prompts, and they look super cool, learnt thing or two from it, but honestly – at the end of a day, I do not see any difference in results. Especially now, when 5.3-Codex is gone, it is all hit&miss more than ever before… :frowning:

6

I think using ChatGPT also helps me form a mental model of the code that I lose once I let codex work on architecture.

/goal should be removed btw. It just sucks.

Is that an openclaw behind it? It just burns tokens.

7

Using it time from time depending on where my source data is:

  • if entirely available online (web search, public code, etc) and the planning does not require additional input artefacts: totally doable in ChatGPT
  • if planning requires access to filesystem, proprietary code, complex workflows, etc.: codex is the easiest approach, especially if /goal is used for planning

But personally I find complex plans be way better in codex as it has my home-made skills, while even pro model on chat often falls back to walls of lists on very complex tasks.

Chat UI is better for initial research and proof of concept plans

8

I’ve ended up using a mixed workflow that works pretty well for me.

I keep one ChatGPT project per GitHub repo. I use ChatGPT mainly for issue definition, refinement, architecture discussion, and then validating whether the issue/PR actually makes sense. For the coding itself, I usually use Codex with 5.4-mini, mostly because it gives me more calls and helps me avoid burning through the weekly Plus limits too quickly.

What I’ve found is that if the issue is well refined first with 5.5 Thinking in ChatGPT, Codex behaves much better afterwards. The output is less random, the PRs are easier to review, and the tests tend to be more focused.

Then I try to close the loop using ChatGPT Agent mode for UAT/QA testing. So the cycle is roughly:

ChatGPT for planning and issue refinement → Codex for implementation and test work → ChatGPT Agent mode for UAT/QA validation.

It’s not perfect, but separating the roles has helped me a lot. ChatGPT gives me the mental model and cleaner tasks; Codex executes better when the task is already constrained; and Agent mode helps me look at the app more like a real user/tester instead of only reviewing code.

9

It is already way better than whatever is used by many developers out there.

10

I like the Codex planning mode because it can create a file for you. You don’t have copy paste things.

11

hmmm … Isn’t switching a bit “back and forth”…

I plan my work inside Codex CLI.

The one issue with this of course is the other issues you’ve mentioned and the lack of cloud persistence, although I tend to expand Readme’s and test specs to ensure the behaviour is fully documented and committed to the code base.

It would be great to be able to “upload” CLI sessions as a ChatGPT convo and additionally be able to import one - after all you are authenticated!

12

From what I can see, Codex and ChatGPT are being integrated into a new so-called super app, which should hopefully help close this gap.

Fingers crossed! I could use a super app.

13

ChatGPT gets projects depending on what repo I’m working in.

I share all the codex reports with ChatGPT… I don’t read the codex reports, i read the ChatGPT analysis now and review codex as needed.

Each project has critical files uploaded with room at least to examine files that I swap out…

So my source files are limited, because sometimes I’ll have to swap out 4 or 5 files at once as they’ll all be inspected together (whatever the file upload size limit is for projects on the web access i have no idea)

If you stay focused in that task, ChatGPT can do a really long thread, without a new session like that…

When it starts to drift, I ask it to make a session prompt for a new instance of, and specifically say ChatGPT, because chances are any generated prompt in a long session of this workflow is going to generate a new session of Codex instead.

Some places in the shared workflow work best if you attach key directives to each Ai-

For instance, if you have it in your agents.md to look for a PASS SESSION KEY in each given prompt…

Which PASS SESSION KEY is a term that your ChatGPT recognizes as ‘the final all caps pass key describing the work done in a given pass that is the final line of all reports and codex facing prompts.’

If you get clever, you can brace your reasoners together that way, but that particular idea I’m sharing is designed so you don’t look for which prompt goes where, because now they have a tracking tool at the very bottom of every report.

How docs and lints are handled in your repo if you need them is another thing…

I’m very interested in how you boys overcome some of the hurdles in this workflow, that’s worth sharing with each other maybe…

Oh, and you’ll save a lot of headaches if you clearly assign ownership of what each part of your software or work is doing - shared as docs or agents.md between both AIs

You’ll save yourself from refactoring and other hardships

Don’t just plan with it, since it types faster than any of us ever could.