Cold Read

The cold-read skill owns the cover-to-cover cold-read protocol — a single reader, reading the whole manuscript in order, carrying a rolling ledger, and logging every defect that survives the per-chapter QA. It exists because a class of defect is invisible to every memoryless pass: a linter sees a line, a chapter reviewer sees a chapter, a continuity query sees a query — none of them accumulate reader-state. Clock drift, cross-chapter state contradictions, prop custody, promise bookkeeping at book scale, knowledge-state slips, and audit coverage bias are all accumulating-memory defects. The cold read is the only pass that catches them.

What this skill is

• A protocol, not a single prompt. It runs on four files (charter, ledger, issues, batch reports) that together make a multi-sitting read behave like one continuous reader.
• Assessment-only by default (rule zero): the read measures; it does not edit the thing it measures. AUTO-FIX is opt-in, narrowly scoped, logged.
• Batched for sustainability and context-window safety: 8–11 chapters per sitting, the ledger the only inter-sitting state.
• A self-contained reference: it restates the severity taxonomy, the issue format, the batching model, and the orchestration pattern so the agent and the muse can run it without loading other skills. Cross-references to gates/, kb-integrity/, and story-ledgers/ are pointers only.

The document set

Load the resource needed

• resources/charter.md — the reader persona ("devoted genre reader who just finished the previous book, with a line-editor's ear; experience first, diagnosis second"), rule zero, the batch plan, the severity taxonomy, AUTO-FIX vs PROPOSE tiers, voice cards, tripwires, PROTECT, deliverables.
• resources/reader-ledger.md — the rolling ledger: NEXT marker with WATCH list, position/clock table, promise register (HOT/WARM/COLD), character knowledge map, prop custody, do-not-re-explain register, PROTECT, tripwires, open threads, grades table. "A stale ledger is an amnesiac reader."
• resources/issue-format.md — the fixed line format CR-### | ch:line | SEV | CAT | "quote" | reader-moment | fix direction | SCOPE, the severity/category vocab, the severity→shared-schema mapping, and the rules (reader-moment mandatory, append-only, continuous numbering).
• resources/batching.md — 8–11 chapter batches cut at natural milestones, the per-sitting resume recipe (charter + ledger + last ~40 issues → continue from NEXT), the per-batch procedure, the batch report contents, the orchestration model, and when to run.

Who runs it

• Muse sets up work/cold-reads/<YYYY-MM>/ from templates/cold-read/ (via /vellum:cold-read), seeds the ledger, and spawns @cold-reader per batch.
• @cold-reader executes one batch: reads charter + ledger + last ~40 issues, continues from NEXT, reads only the batch's chapters, returns (a) append-ready issue lines, (b) ledger section updates, (c) the batch report. Assessment-only; no peeking at outline or editorial notes during a batch.
• For 15+ chapters muse MAY use the digest pattern from fiction/agents/review-coordinator.md (MIT) — spawn per-batch readers, aggregate from summaries, cache per-batch reports on disk.

How it fits the gates

Cold-read findings are advisory by default; BLOCKER-class findings block export until resolved or dismissed-with-author-signoff (the same discipline as story-ledgers/ exemptions). MAJOR/MODERATE need a triage note. The export readiness checklist (gates/resources/readiness.md) cross-checks the cold-read issues.md for open BLOCKERs.