When we shipped AI agents as an automation task type, a skill was something you picked from a dropdown and hoped for the best. You could see its name and one line of description. You couldn't see what it actually told the agent to do, you couldn't change it without an admin, and a skill written in one project was invisible in every other one.
This release fixes all of that. A skill is now a product object you own.
See it before you run it
Selecting a skill renders its full instructions, read-only, right in the task config. That's the whole system prompt — what it looks at, how it reasons, what it must never do. No more running something you can't read.
Name your own copy
Databasin's own skills — executive summary, data-quality summary, anomaly and root-cause explainer — are gold copies. Picking one always makes you a copy, and you name it something meaningful before anything is created. Your library stays yours; ours stays pristine. And your copy keeps its lineage, so when we improve the original you'll see "update available" and decide whether to take it.
Edit it where you use it
A skill you customize inside an automation belongs to that automation. Edit its instructions, its description, and the inputs it asks for — right there in the task, without opening another screen and without changing behavior for anybody else. Every save is a new version.
Choose who else gets it
Three scopes, and widening is one click:
- Just this automation — your copy, your tuning, invisible in everyone else's picker
- This project — the team on this data can run it and fork their own
- Your whole organization — every project sees it
Skills authored in the library default to org-wide, because that's how most teams actually think about them. And anyone who can work in a project can write one — no admin ticket. Admins keep an org-wide governance view of everything that's been built, including the private per-automation tuning, with one-click promotion for the ones worth sharing.
Skills now declare their own inputs
Previously every skill asked for the same three things — a table, some starter SQL, a prompt — whether it needed them or not. Now a skill declares its own fields, and the task form is generated from them: table pickers, dropdowns, numbers, free text, whatever it actually needs. Multi-table selections now reach the agent intact.
Same guardrails, same blast radius
None of this widens what an agent can do. Still three read-only tools, still SELECT-only on every query, still your turn limit and query budget, still a full audit trail per run — and the data connection is chosen by the task, never by the skill. A skill anyone can write is still a skill that can only read.