What problem does it solve? Writing accurate commit messages, PR descriptions, and squash messages is error-prone: agents invent test results, guess issue links, ignore the repository's own conventions, and leak AI attribution trailers into history. This Skill turns verified repository evidence into proposed change-message text while enforcing a strict output-only boundary. ## Core Features & Use Cases - Evidence-based authoring: Every claim is routed through a six-step evidence ladder (verified diff, gate results, repo policy, observed history, linked work items, generic fallback); unsupported claims are omitted or marked with a TODO placeholder. - Four message modes: Commit messages, pull-request title/body scoped to the merge-base footprint, squash messages that synthesize the final delta, and audits of existing draft messages. - Convention and attribution policy: Detects and preserves repo-native commit conventions (Conventional Commits is only a fallback) and default-prohibits model/tool attribution trailers, generated-by footers, and badges unless explicitly requested. - Use Case: After finishing a feature branch, ask for a PR description; the Skill reads the verified diff and actual test output, follows the repo's PR template, and proposes title and body text — without ever running git or gh commands itself. ## Quick Start Ask the agent to draft a commit message or PR body for the current verified diff, and it will propose text only while you retain control of committing and publishing.
