What problem does it solve? When you hand-edit AI-generated drafts, those corrections contain valuable signal about your voice and preferences, but that signal is normally lost. This Skill captures snapshots of generated text at each pipeline stage, diffs them against your manual edits, and turns recurring patterns into tested improvements to your writing registers, skill rules, and review agents. ## Core Features & Use Cases - Snapshot capture: Saves post-review, post-fixes, brief, review-findings, and suppression-log snapshots with a manifest so every pipeline stage can be diffed later. - Learning analysis: Dispatches a learn-review agent over a minibatch of edited pieces to propose tiered candidate edits (Apply / Hold / Reinforce / Contradictions) to registers, rules, or agents. - Gated edit acceptance: Discipline edits pass an objective script check; taste edits pass a human A/B pairwise gate with a shadow judge, so only validated changes land. - Evaluator correction mode: A separate loop tunes the prose-review and craft-review agents using your accept/reject decisions as ground truth, with precision and recall metrics per agent. - Use Case: After editing a generated blog post by hand, run the learning analysis to discover you consistently remove a certain construction; the Skill proposes a register rule, gates it against held-out pieces, and retains the winning before/after pair as a demonstrated edit. ## Quick Start After manually editing a piece generated with copydesk, ask Claude to run /copydesk:learn to analyze your edits and propose improvements to your register and review rules.
