@edwinhu
The phase-gate pattern in writing-revise is doing something most prose-revision tools skip: it blocks edits until REVIEW.md exists, so the revision loop always starts from structured diagnostics instead of freeform rewriting. That constraint is what makes the loop converge -- critical/major issues drive the iteration, minors stay advisory, and the geometric-mean rubric forces fixes toward the weakest dimension rather than over-polishing what already scores well.
I set up an online entry point so anyone reading the repo can try the revision workflow directly on their own draft:
A visitor uploads a draft, the skill runs the review-then-revise loop, and they get back structured feedback with targeted fixes -- the same transactional-save discipline your transactional_fix.py enforces locally. They can run the revision workflow here.
Every session leaves a usage record. The questions people actually type, the drafts they upload, where they retry -- that signal is useful for sharpening SKILL.md and the rhythm rubric against real revision patterns beyond the validation run documented in rhythm-lessons.md.
Feel free to close if this isn't relevant.
shesonglin@tinkerland.ai
@edwinhu
The phase-gate pattern in writing-revise is doing something most prose-revision tools skip: it blocks edits until REVIEW.md exists, so the revision loop always starts from structured diagnostics instead of freeform rewriting. That constraint is what makes the loop converge -- critical/major issues drive the iteration, minors stay advisory, and the geometric-mean rubric forces fixes toward the weakest dimension rather than over-polishing what already scores well.
I set up an online entry point so anyone reading the repo can try the revision workflow directly on their own draft:
A visitor uploads a draft, the skill runs the review-then-revise loop, and they get back structured feedback with targeted fixes -- the same transactional-save discipline your
transactional_fix.pyenforces locally. They can run the revision workflow here.Every session leaves a usage record. The questions people actually type, the drafts they upload, where they retry -- that signal is useful for sharpening SKILL.md and the rhythm rubric against real revision patterns beyond the validation run documented in rhythm-lessons.md.
Feel free to close if this isn't relevant.
shesonglin@tinkerland.ai