you are not codexmaxxing enough and this is gonna be a long one so sit tight.
before i get into the weeds of codexmaxxing, i need a mindset shift from you so just repeat this in your brain a few times
"i don't know if codex will be able to do it, let me try it"
this is important and i'll tell you why it is. past few weeks, i've been trying to improve my workflows around coding and building things and almost always the bottleneck is my mindset
I JUST NEED TO BE ABLE TO GIVE CODEX HARDER TASKS AND TRUST THAT IT WILL GET IT DONE.
this has been my number 1 realisation over the past 2 weeks. has prolly helped me more than anything else you will read below.
now let's jump right to it:
we will start with an easy one but stay till the end, it keeps getting better. the last one is my favorite(no this isn't to make you read till the end, it actually is)
/goal
before you hit that back button, stay w me.
everyone already knows /goal is incredible but here are some pitfalls you must avoid:
- giving it permission to do everything. this wasn't me. this was codex trying to rank me 1 for SEO and it assumed that it had the permissions to push code through to main. my bad to assume that it would know better and not explicitly saying you can only create PRs.
- not making the progress verifiable. if there is an mcp that could track the progress, add it. if there is an api you could call, mention it. give it everything to make the progress verifiable. you should see progress in the metrics, then and only then is it gonna create anything valuable.
- not defining what's done. "improve this website" okay. improve it to what? there should be an end goal or it could run around in circles burning through tokens far exceeding what you thought of. a useful definition of done could be a test suite passing or reaching a benchmark or anything that can be quantified really.
you can also pause a goal, edit it at any time btw.
there are 3 different browsers in codex
no way right?
- $browser : for the in-app browser in the side panel, where Codex can inspect and annotate web surfaces
- @chrome : for signed-in browser state and Chrome-based workflows
- @computer : for work that only exists through a desktop GUI
thank you @jxnlco for this.
the default is $browser, that's the side panel window that's already in codex. you can use chrome for signed in browser work but also not many people know this that you can provide credentials in an env file in .codex that could do the thing just well and faster.
steering and /side
you can queue messages, that's okay but you can also steer a conversation. i like to look through long running chats to see the direction and often write a message and then click on steer to give it right instructions.
it could be anything from:
- ayo why did you create a new row for this, just add it in the jsonb
to
- ayo why did you add it in the jsonb, just create a new row for this
you get my point.
/side is p useful too, what that creates is a sidebar chat window to chat w codex while your agent is running. i often ask it implementation details, decisions, assumptions while its completing a feature request.
one v cool fact about /side is that it can let you review and change markdown files that /goal or plan mode created without stopping the implementation.
@mattpocockuk's skills
i added a new section for this because it is that good.
i've been a user of /grill-me for a while now, earlier it was ask_user_question_tool in claude code and now its just grill me. the updated version of grill me is /grill-with-docs which is even more impressive.
he also talks about best practices to create a skill:
- Concise
- Responsible for one thing, not multi-step
- Composable
- Progressively disclosed
- Harness-agnostic
one thing i did wrong p early on was i created my own skill for planning and that had writing the plan and grill me in the same skill and it was always eager to run to the second stage before completing the first one.
i've also learnt about the dumb zone from him basically anything that crosses 100k context window, its always better to start a new chat when that happens or compact the chat.
i don't really know the context that stays after compaction so i prefer creating a markdown and moving it over to a new chat.
voice mode
i know it feels weird.
i know you can type p fast.
but did you know you could speak faster?
i tried wisprflow a few months back(this is not an advertisement and you should use whatever you like) but i've since moved to voice first codexmaxxing especially for the first prompt.
there is something about speaking that makes you talk about nuances, give more details when compared to typing. i guess its just easier but yeah.
hooks
oh man, not again. not another post about hooks.
i'll keep it short.
you already know that hook is a process that can run after or before a chat process completes.
the interesting part is before.
if you have used codex enough, you already know how worktrees work. but did you know you could configure a pre-hook to worktree creation so that it uses the common node modules and env from .codex?
yeah me neither. the point is, if there is anything you always do pre or post a process, its almost always better to configure a hook for it. just ask codex to do it for you.
github actions
i know github ain't the most dependable piece of software but it does give you a fair share of tools to do things.
two pretty cool usecases we have built at @lucent_ai:
1. sentry to issue to pr creation: there is a codex cloud agent running that we run to fix issues that are up in sentry. turns out if you give it a good enough prompt and verifiability, it runs it as good as codex.(omg who knew)
2. pr review fixes: this is a fun one.
so we have created a github bot that looks at every PR comment and fixes/reviews the comment. it took a few turns but its incredible because i literally don't have to look and fix claude-code review comments anymore.
so now @sama works for me as an intern which is pretty cool if you ask me.
there are also a few noteworthy mentions which i haven't either gone v deep in or everyone knows about them
- codex on mobile
- automations
now if you are here till the end, i know this might be overwhelming but i try something new everyday and remind myself what i said at the start:
I JUST NEED TO BE ABLE TO GIVE CODEX HARDER TASKS AND TRUST THAT IT WILL GET IT DONE.
and that mostly works. this ain't the perfect guide so feel free to add things in the comments. i appreciate you so much for reading this till the end. bookmark this, rawdog it rn, do whatever but codexmaxxxxx
thank you to a lot of people who have taught me this. forever in debt :)
rt if you want your twitter fam to codexmax as well.
see ya and have a great day my ai agents.