What problem does it solve? ADR draft files accumulate in docs/adr/ faster than the automatic Stop-hook threshold triggers cleanup, and LLM refusals or min-interval locks can delay normalization. This Skill lets you manually run one normalization batch on demand, bypassing the threshold and interval guards, and review the result via git diff before committing. ## Core Features & Use Cases - Manual batch execution: Runs .claude/hooks/normalize-decisions.sh with ADR_NORMALIZE_THRESHOLD=0 and ADR_NORMALIZE_MIN_INTERVAL_HOURS=0 to force a single batch through the LLM. - Dry-run preview: ADR_NORMALIZE_DRY=1 writes the planned operations to the log without touching any files. - Review and rollback workflow: Inspect delete/rewrite/merge-into operations with git status and git diff, then accept, selectively revert, or clean the whole batch. - Worktree isolation: Enforces a preflight that creates a dedicated git worktree (<current-branch>-adr-normal) so normalization never runs in the main tree or in parallel. - Use Case: You have 25 ADR drafts from recent sessions and want to clean the inbox now instead of waiting for the threshold — run the dry-run, check the log, execute the real batch, review the diff, and commit. ## Quick Start Ask the agent to run the n-adr-normalize skill to process one batch of ADR drafts in docs/adr/ and show the git diff of the results.