For creators
Write a meaningful skill update—and know when to roll back
Use a small release record and a repeatable test to make changes easier to trust.
Quick answer
Prepare updates as new candidates and test the behavior you intend to change. The approved version stays available while a candidate is reviewed. If a release introduces a problem, version history can restore an eligible previously approved version after a fresh scan; rejected or unreviewed text is not a rollback target.
Define the change with a before-and-after case
Suppose a fictional meeting-summary skill produces decisions and action items. Users report that it turns tentative suggestions into commitments. The update should distinguish an agreed action from a proposal, rather than merely ask the AI to be more accurate. Keep the output fields familiar so the buyer can see what changed.
Create one test note containing both “Mira will send the draft Friday” and “perhaps we should ask sales.” The expected result has one assigned action and one unresolved suggestion. This concrete example is a better release check than comparing how polished two summaries sound.
Save a candidate without replacing the release
Open the existing skill in the creator editor, revise the instructions and relevant description, and save a draft while testing. Submit the finished candidate for scanning and review. Submitted, scanning, and in-review states describe the candidate; they do not mean buyers are already receiving that text.
The previously approved release remains available during review, unless a separate safety or access restriction applies. Do not advertise the tentative-suggestion fix as delivered while it is still pending. Keep a short local change note with the problem, changed rule, example input, and expected output.
Check what a new conversation receives
After approval, begin a fresh test and ask the AI to load the full skill. Confirm the title and inspect whether the revised rule changes the example as intended. An old conversation may still contain the earlier instructions; publishing a release does not rewrite text already delivered there.
Long loads can arrive in several parts. Skillset keeps those parts tied to the version selected for that load, so an update should not splice one release’s opening steps into another release’s ending. Start a new load when deliberately testing the newest approved version.
Use rollback for a known approved alternative
If the new release loses an essential output field, inspect version history and select a previously approved clean version that contained it. Restoration runs the current scan again. A version that now raises a safety concern needs review instead of automatic restoration.
Rollback restores the selected version’s saved listing metadata as well as its instructions. Check title, description, dependencies, and displayed pricing before using it as a repair. It does not promise to rewrite existing payment agreements. Record why you restored the release, then prepare a corrected candidate using the failing test.
Common questions
Can I roll back to my last saved draft?
No. A draft has not earned approval. Rollback is reserved for an eligible previously approved clean version, and the current scan must still allow it.
Will buyers lose access while an update is reviewed?
A pending candidate does not replace the approved release. Independent safety restrictions or changes to a buyer’s entitlement can still affect access.
Do I need to retest a tiny wording change?
Test any edit that changes a decision, output, permission, or dependency. A short sentence can materially alter behavior even when the document barely changes.
Put a workflow to work.
Connect your library to your AI, or turn your method into a skill.