Allow specifying `model` type to be used in a skill
Resolved 💬 4 comments Opened Feb 8, 2026 by guidedways Closed Feb 8, 2026
💡 Likely answer: A maintainer (miraclebakelaser, contributor)
responded on this thread — see the highlighted reply below.
What variant of Codex are you using?
CLI
What feature would you like to see?
I've got a $localize skill that I use for localizations. I've noticed 5.3-codex-xhigh is just very bad at copy - it writes weirdly, robotic sentences that resemble commentary within code. 5.2-high writes beautifully. I'd like to be able to specify the model to use for this skill.
Additional information
_No response_
4 Comments
Directionally, I agree with you (in my case, I don't need xhigh spending minutes committing changes to git). But I don't know if attaching models to skills is the right way to do it.
Given that a skill isn't a context per se, but essentially lazy-loaded instructions that persist in the session context, it seems messy to map models to skills. It invites unpredictable behavior. For instance, what if the agent invokes multiple skills (that have different model specs) in the same turn?
Instead of attaching models to skills, what do you think about attaching models to definable units of work?
| Attaching a model to a... | ..results in the model being used for.. |
|----------------------------|------------------------------|
|
/review(already exists) | the entirety of the review || custom
/prompt| the full agent turn directly after the/prompt, before the next user turn || custom subagent | the entirety of the subagent's session |
I don't think it makes sense for individual skills to specify a model. Skills can be nested (refer to each other) and combined with each other. They would no longer be composable if models were specified as part of the skill.
@etraut-openai respectfully, that's the idea. Since a skill can refer to another skill, this allows nesting skills _and_ using a specific model for a specific skill. For instance
$documentationcan use codex-spark, while$my-special-reviewcan usegpt-5.2-xhighand$implement-plancan use gpt-5.3-codexI've found gpt-5.3-codex is incredible at coding and reasoning, but not as good (read: very bad) with
$localizeand$write-user-guidestyle skills - it localizes strings as if it's writing documentation. I end up switching models manually between skills and I'm unable to nest skills due to this.@guidedways have you tried using custom agents for this? You could configure a localization agent with custom developer instructions. That agent could use your localization skill.
I haven't played with it much yet, but I have set the explorer agent to use
gpt-5.3-codex-spark high– it's working well.