What problem does it solve? Instruction assets like AGENTS.md, rules, skills, and prompts grow endlessly because nobody deletes sentences, until the model attends to none of them. This Skill applies a scheduled deletion discipline that cuts dead weight and lands changes where a person can review them file by file. ## Core Features & Use Cases - Sentence-level audit: Walks every instruction file and classifies findings as no-op, duplicate, prohibition, slop, dead-pointer, or over-budget, with a cut, rewrite, or report action per hit. - Review-safe delivery: Opens one draft pull request with one commit per file and a ledger of every cut, so a reviewer can drop a single file without losing the rest. - Rejection memory: Skips any file touched by a previously rejected instruction-review pull request, so the run converges instead of re-proposing refused cuts. - Use Case: Run weekly against a repository whose AGENTS.md and skill files have bloated over months; receive a draft PR trimming redundant rules and fixing dead pointers, with protected frontmatter and managed sections untouched. ## Quick Start Ask the agent to run the scheduled instruction review with default scope and rewrites enabled, then review the resulting draft pull request.
