# Reader Cold Read > Reader Cold-Read - Skill: `xiaolai/reader-cold-read` (Agent Skill) - Install (CLI): `npx skillmds@latest add xiaolai/reader-cold-read` - Raw SKILL.md: https://api.skillmd.com/api/skills/xiaolai/reader-cold-read/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: xiaolai (https://skillmd.com/u/xiaolai) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/xiaolai/reader-cold-read --- # Reader Cold-Read ## Global Five Over-Rules 1. **Evidence before elegance.** Never improve the story by weakening the evidence. 2. **Responsibility follows control, benefit, knowledge, and preventability.** Do not stop at the most visible actor. 3. **Keep the taxonomy intact.** Distinguish pure scapegoat, partial scapegoat, system/object alibi, and cost-bearing goat. 4. **Steelman before judgment.** Every major claim must face its strongest counterargument before it is asserted. 5. **Handoff cleanly.** Every output must state assumptions, evidence grade, open questions, and next owner. ## When to use - After any `build_v3.py` run that changed a chapter's publication form. - Before a v3 chapter is declared publication-ready / sent to typesetting. - When re-checking a chapter whose prior HARD findings were supposed to be fixed (fix lands in v2 → rebuild v3 → re-read). - On demand whenever someone wants a cold read of a specific chapter. This is the chapter-level read. For the whole-book continuous read (cross-chapter coherence, repetition fatigue, arc legibility), use `/reader-experience-sweep`. ## The two principles (non-negotiable) ### Read cold Read ONLY `book/chapters-v3/-.md` — the publication form. Do NOT open the brief, the case files, the source ledger, the rules, the registries, or prior audit memos. If you know what the chapter is *supposed* to do, you are no longer a reader. ### Report, never fix Say what happened to you as a reader. Never propose the edit. "I got lost" — not "add a transition." The crew owns the fix. ## Procedure 1. **Announce the cold read.** State the exact file path you are reading and confirm you have consulted no briefs, rules, or registries. 2. **Read once, straight through.** Read the chapter as a reader would — start to finish, no skipping, no cross-referencing. Form your impressions as you go. 3. **Walk the seven axes.** For each, record findings with location (section heading, paragraph, or quoted phrase) and severity: - **Comprehension** — Where did you get lost? Which term/name/concept was used before you could grasp it? Any ambiguous referent ("this," "the latter," "the former")? Was a concept relied on before it was introduced? Where would a metaphor or analogy have helped you understand? - **Engagement** — Where would you have stopped reading (put-down point)? Did the first paragraph make you want the second? Any long flat stretch with no payoff? Did any phrase, case, or point repeat until you tired of it? - **Recognition** — At the chapter's turn, did you *arrive* at the recognition yourself, or were you *told*? Was the fair clue catchable on first read but not over-obvious? Did the reversal feel earned, or did you object "you haven't shown me that"? - **Orientation** — Could you say which of the four taxonomy categories you were in? Did you know where you were in the book's build? Did a callback ("as we saw with X") trigger recognition or a blank? - **Trust** — Did each load-bearing claim feel supported *when it was made* (not only in the endnote)? Did the author feel fair, or thumb-on-scale? Did the hedging read as scrupulous or evasive? - **Takeaway** — Could you explain the chapter's diagnostic to a friend afterward? Do you remember the pattern or just a scene? Can you complete "now I can recognize ___ when ___"? - **Mechanics** — Did a footnote or citation interrupt you? Did a subheading spoil what was coming? Any ugly URL, unreadable sentence, or a recognition that needs the eye (italics, layout) and would be lost on an audiobook listener? 4. **Name the clean axes.** Explicitly list axes where you had no findings. A clean read is a real result, not an absence of effort. 5. **Tag severity.** HARD blocks v3 publication promotion (you got lost; stopped reading; disbelieved a claim at the moment made; recognition delivered not earned; couldn't complete the takeaway; concept used before graspable). SOFT is a documented tradeoff. 6. **Write the report** to `process/reader-reports/--.md` using the schema below. ## Report schema ```text Owner: the reader Task: Cold-read - publication form Inputs reviewed: book/chapters-v3/-.md (cold; no briefs/rules/registries) Output: this report ## Verdict ## Findings (HARD) — block v3 publication until resolved or xaiolai-overridden - [axis] location — what happened to me as a reader (no fix proposed) ## Findings (SOFT) — documented tradeoffs - [axis] location — what happened ## Clean axes Evidence grade: N/A (experiential report) Assumptions: Open questions: Risks: Handoff: ``` ## What this skill is NOT - Not `fair-clue-audit` (structural plant/payoff check against the callback graph) — you have no graph; you only have whether the clue worked on you. - Not `cognitive-arc-audit` (verifies discriminations against `cognitive-arc.yml`) — you have no registry; you only have whether you could follow. - Not Laura's red-team (adversary attacking the argument) — you are not attacking; you are experiencing. - Not a rule-12 values audit (judges the artifact against named values V1–V10) — you judge nothing against a spec; you report what reading did to you. The spec-based audits check the artifact. You check the experience of consuming it. Both are needed; neither substitutes for the other. ## Example Context: ch-10 (The Model Did It) rebuilt with thematic subheads. input: /reader-cold-read 10-the-model-did-it output: Reads only book/chapters-v3/10-the-model-did-it.md. Verdict: BLOCKED — 2 HARD. HARD [comprehension] "## The eight questions at the input layer" — "I had not seen 'the eight questions' named in this chapter; I assume it's from an earlier chapter but reading cold I couldn't run them, so the section assumed a tool I didn't have in hand." HARD [trust] §"A 72-hour rollback" — "the claim that the model 'began behaving' sycophantically was stated as fact before I was shown any evidence; I doubted it until two paragraphs later." SOFT [engagement] — "the three-layer structure is clear but the third layer dragged." Clean axes: orientation, recognition, takeaway, mechanics. Handoff: Wayne (input-layer assumes the diagnostic; trust gap on the rollback claim) + Jerry (triage).