rule-miner

# Rule Miner Turn the corrections you keep making into durable rules - and keep the rules you already wrote honest. It runs as a **dynamic [Workflow](#how-it-runs)**: separate agents with their own context windows, so verification can't be biased by the agent that proposed the rule. ## What this skill is - and isn't | Skill | Question it answers | Output | |---|---|---| | **`vd:rule-miner` (--mine)** | **"What do I keep correcting that should be a rule?"** | **CLAUDE.md rule proposals, evidence-backed** | | **`vd:rule-miner` (--check)** | **"Do my existing rules still hold on this code?"** | **Confirmed rule violations, false-positives filtered** | | `vd:docs` | "Is the team documentation current?" | Updated `docs/` | | `vd:code-review` | "Is this change ready to land?" | Inline PR comments | Rule-miner manages the **rules Claude follows** (CLAUDE.md, `~/.claude/rules/*.md`), not project docs and not a specific diff's review. It proposes; it never silently rewrites your rule files. ## Two directions | Mode | Direction | What it does | |---|---|---| | `--mine` *(default)* | history → rules | Sweep sessions + git + PR reviews for repeated corrections, cluster, verify each candidate, distill survivors into proposed rules | | `--check` | rules → code | Load the existing rule set, run one verifier agent per rule against a diff/codebase, skeptic-filter false positives, report real violations | Detect from the argument. "what do I keep correcting / turn my feedback into rules...