> Very branchy. Very un-DRY.
replyI've found this can be vastly reduced with AGENTS.md instructions, at least with codex/gpt-5.4.
What sorts of instructions?
replyUsually I just put something like "Prefer DRY code". I like to keep my AGENTS.md DRY too :)
replyalso add "no hallucinations" and "make it works this time pretty please" while also say Claude will go to jail if does not do it right should work all the time (so like 60%)
replyThere are of course limits to what prompting can do, but it does steer the models.
replyIn TFA they found that prompting mitigates over-editing up to about 10 percentage points.