If a product can be built from a phone, the harder question becomes what still needs a founder's judgment.
The answer is usually not typing speed.
It is choosing the smallest useful version and knowing when the result is good enough to ship.
The useful part of an AI coding workflow is not asking it to write more code.
It is forcing the work into one issue, one isolated workspace, and one reviewable diff.
That makes speed easier to inspect instead of merely easier to celebrate.
i wrote a small bash script for my linear → codex workflow.
it can:
find the next unblocked linear issue
create a separate git worktree for it
build a scoped codex prompt from the issue
run codex on that issue only
let me review and test the diff
commit it
post the result back to linear
optionally mark the issue done
the part i like most is the isolation:
one issue.
one worktree.
one codex run.
much easier to control than giving an agent a huge feature and hoping it stays in scope.
thinking of publishing the script if anyone wants it.
here is the script
https://t.co/7HCjd2Y3n5
@denisyurchak@yadaphone i am just starting on x. how does x show your launch post to that many people? I see that you got a lot of followers, but should have been low when you made those posts right?
Giving Codex an entire feature is usually a mistake.
Current workflow of my side projects:
ChatGPT -> break the feature into implementation-sized tasks
Linear -> one issue = one clear unit of work
Codex -> solve exactly that issue, test it, commit it
Then repeat.
AI coding gets much more reliable when the agent works through a ticket queue instead of being told to “build the feature.”
Curious how you are structuring your AI coding workflow.
@GarvSanwariya mostly in the ticket. i try to make the issue self contained with scope, expected behavior, edge cases, and what tests should pass.
but sometimes on the fly too lol.