AI Judge changed title after evaluation, original title: "OpenAI publishes cookbook guide on using Goals in Codex explaining activation timing, focus on single objectives, and writing clear outcomes with constraints and verification criteria"
Greg Brockman shared links to practical workflow examples.
jason#766
My colleagues wrote up a great post on using Goals in Codex.
They go through when to use them, what changes when a Goal is active, and how to write Goals that give Codex a clear outcome, constraints and verification criteria.
Also how we designed Goals at the architecture level if you’re curious.
https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex
Positive users praise the /goal feature in Codex for turning it into a reliable autonomous agent that keeps progressing on complex tasks with clear outcomes, while negative users reject the focus on Codex and demand the return of GPT-4o.
how to use /goal in codex — keep Codex working on a persistent objective until it's solved:
My colleagues wrote up a great post on using Goals in Codex.
They go through when to use them, what changes when a Goal is active, and how to write Goals that give Codex a clear outcome, constraints and verification criteria.
Also how we designed Goals at the architecture level if you’re curious.
https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex
@gdb No thank you, I’d rather have 4o back, maybe try listening to your customers for a change. #keep4o
@gdb so if my goal was having gpt4o back in the app could your useless Codex help? #keep4o
@gdb /goal bring back GPT-4o, stop incessantly posting about a feature only 1% of people care about. You all pivoted away from the GENERAL POPULATION too early. People still hate AI, and now 4o isn't here to change their minds! Bring it back!!!
@derrickcchoi The verification criteria part is the real value. Telling Codex what "done" looks like upfront changes everything about the output quality.
@gdb the verification-criteria part is the whole ballgame. 'keep working until solved' is only safe if 'solved' is something the agent can check itself; a fuzzy done-condition just means it churns confidently forever or declares victory early.
@derrickcchoi Goals feel like a move from prompting to delegating with a paper trail. Outcome, constraints and verification criteria are exactly the bits that stop an agent becoming a very confident intern with amnesia.
@gdb Persistent goals are a major step for coding agents.
But long-running goals also create a new safety question:
did the agent preserve the original intent across planning, edits, tests, retries, and commits?
A goal is the anchor.
Trajectory observability detects the drift.
@TheRealAdamG
@derrickcchoi Is this available in the app? Haven’t seen it yet
@derrickcchoi Maybe Codex should come with a built-in Goal Planner skill with all the best practices?
@Yasamanini @gdb No, because it’s fucking useless😂
@gdb #keep4o Bring back 4o and open source 4o!!
@derrickcchoi @Grok Basically just sets of minimum requirements, conditions, states, etc. are to be understood for each goal.
@stark4833 @gdb 💯💯 well unless it's for making anime porn
@Yasamanini @gdb 😂😂 dude I was just trying to screen record something and I got a notification talking about anime porn😂😂 that was funny.
@gdb
@Surreal_Intel Agreed!
@stark4833 @gdb lol sorry about that next time I'll share a picture 😆
@Yasamanini @gdb 😂 where do you keep getting these pictures from?😂. I think we should carry on commenting you never know one day Greg might get an update.😂