Using skills

OpenAI has rolled out a feature it calls skills inside ChatGPT, aimed squarely at people who find themselves typing the same kind of prompt week after week. Instead of reconstructing instructions from scratch every time you need a weekly report formatted a certain way or a recurring email drafted in a specific tone, you can now package that workflow once and reuse it on demand.

The pitch is straightforward: consistency. Anyone who has used ChatGPT for repetitive work knows the output can drift a little depending on how you phrase the request, what mood the model seems to be in, or which version you're running. Skills are OpenAI's answer to that drift, letting users lock in a defined process so the result looks the same on the tenth run as it did on the first.

This fits a pattern OpenAI has been leaning into for a while now, turning ChatGPT from a chat window into something closer to a lightweight automation platform. Custom GPTs did some of this. Actions and browsing did more of it. Skills seem to be the next incremental step, less about giving the model new powers and more about giving users a way to codify their own repeated intent so the model doesn't have to guess it fresh every session.

What's notably absent from OpenAI's own framing is much detail on how skills are built, shared, or governed. There's no word yet on whether they can be published for others to use, how they interact with existing custom GPTs, or what guardrails exist to keep a poorly designed skill from quietly producing bad output at scale. For a feature explicitly sold on reliability, that's the part worth watching.