review-to-lintlisted

A procedure for extracting the mechanically decidable items from an existing reviewer's (agent / review skill) checklist, or from past sessions' reviewer history, into a deterministic script, thinning the reviewer down to semantic checks only. Use via /review-to-lint when the author says "turn this reviewer into a lint", "absorb the review into a lint", "move the mechanical checks into a script", "find what in the history can be linted", or when a reviewer's findings show repeated mechanical items. NOT for — creating a new reviewer (→ skill-creator), changing the semantic criteria themselves (each reviewer owns them), designing a judge (→ llm-as-judge), re-extracting reviewers that already have an evidence script (done for readme-writer / adr-writer).
shimo4228/claude-harness · ★ 3 · Code & Development · score 79

Install: claude install-skill shimo4228/claude-harness

# Review → Lint absorption A reviewer's checklist mixes in items that a script counts more accurately and cheaply than an LLM does. This skill draws that boundary, moves the mechanical side down into a script, and focuses the reviewer's attention on semantic checks. Prior examples: `readme_evidence.py` (readme-writer), `adr_lint.py` (adr-writer, ADR-0051). The division-of-labor principle is "existence = code, content = LLM" (introduced by ADR-0021, applied to ADRs by ADR-0044) and feedback: deterministic_semantic_layering (script measures + LLM interprets). ## 0. Entry — checklist-first vs history-first If the target reviewer is already decided, go straight to §1. If you are deciding "which reviewer / which convention should be linted," mine the reviewers' history. A checklist-first inventory surfaces only the items a reviewer **holds in writing**, so it misses findings that actually recur (measured 2026-08-29: RFC-0005's 12 candidates covered only the document / skill-asset layer and missed 3 classes targeting `hooks/*.sh` = #13–#15 of the same RFC). **Build no mechanism** — create neither an extraction script nor a collection hook (ADR-0055 Decision 5, "build no collection mechanism"). One manual investigation is enough. If a second one is requested, propose a script at that time, including a supersede of ADR-0055. Non-obvious points about the corpus and extraction: - Reviewer reports are at `~/.claude/projects/<project>/<session-id>/subagents/agent-*.jsonl`. **Not