An AI-generated revision should arrive as a candidate, not a silent replacement. SkillForge binds each proposal to the target revision that produced it, shows which sections changed, and leaves the source untouched until you decide what belongs.
Check the proposal is still current
Confirm that the target has not changed since the request began. A proposal based on an older revision can overwrite newer reasoning even when its text looks useful. SkillForge rejects stale proposals and desktop saves also check whether the underlying file changed outside the application.
Read the summary as a claim
The proposal summary should name the actual improvement. “Made it better” is not reviewable. Look for a concrete statement such as “added an evidence priority and a stop condition for an unavailable deployment marker.” Then verify the changed text does that and nothing broader.
Trace changes to the request
Use the selected reference passage, your annotation, and any additional feedback as the acceptance basis. A change without that support may still be good, but it needs its own reason. Reject copied branding, repository-specific instructions, tool grants, or permissions that do not fit the target.
Accept at the section boundary
Review one semantic section at a time. Accept a section when it improves the intended behavior and preserves unrelated content. Edit it when the mechanism is right but the scope or wording is wrong. Reject it when it solves another problem or introduces an unsupported claim.
Inspect the resulting whole
Section-level review can still create repetition or contradiction across the document. Read the full skill after applying chosen changes. Check that trigger language matches the procedure, inputs exist before they are used, permission rules do not conflict, and the definition of done reflects the actual artifact.
Test before saving over a real file
Use fixed prompts that exercise the changed behavior. Compare with the previous revision when the difference is not obvious. For a deployment skill, run the target-specific verification it requires. For a writing skill, compare factual accuracy, structure, and usefulness. A valid parser result cannot prove those outcomes.
Keep a recovery path
Desktop saves use an atomic replacement and create a recovery snapshot manifest. Browser drafts remain in the signed-in user’s current browser and can be downloaded as ordinary SKILL.md. Keep a version you can restore before replacing important shared instructions.
The aim is a visible chain from evidence to intent, proposal, test, and save. That chain gives a later reviewer enough context to keep, revise, or retire the pattern without trusting the original author’s memory.