Make it an owned skill
Outcome
You can package a working tool so someone else — or future you — can run it, trust it, and keep it alive, and you understand why that last part is the whole point.
Concept
A tool that lives only in your head dies when you forget how it worked, or when the model updates and quietly starts giving worse answers. The thing that lasts is a skill: a small package holding the tool itself, the examples that define correct, the eval that measures it, and a note saying who owns it and when to re-check it. Make a folder — the script, your gold set, your eval, and a short README: what it does, how to run it, how to run the eval, who owns it, and the rule "re-run the eval whenever the model changes." That README is what turns a trick into a skill. Naming an owner feels like overkill for a personal project, but it is the single habit that separates AI work that lasts from the pile of demos nobody maintains. When a new model ships, the owner re-runs the eval. If the number held, good. If it dropped, the gold set shows you exactly what broke.
Weak approach vs. strong approach
The ask
You keep your summarizer script in your downloads folder. It worked great in March. A new model version ships quietly in the background, and you keep using the tool the same way you always have.
What comes back
The tool still runs. No errors. You have no idea the summaries are now skipping the action items that used to appear reliably at the end. You shipped three meeting recaps this week with that gap. You won't notice until someone asks why the follow-ups weren't captured — and by then you won't remember whether it was the tool, the model, or just a fluke.
Try it
Package your tool. A folder with the script, your gold examples, your eval, and a README covering what it does, how to run it, how to run the eval, who owns it, and when to re-check. Then pretend a new model just shipped: re-run your eval and see if the number holds.
What to look for: Notice that this is the first version of your work someone else could actually trust — because it carries its own proof and its own maintenance plan.
Takeaway
A skill is a tool that carries its own definition of correct, its own proof, and its own owner. That is what survives a model update — and what "going beyond prompting" actually means.