Horror Vanguard Reviewer, Rebuild Prompt

Rebuild the /horror-vanguard-reviewer skill. Its output is competent and dead. It reads like an essay with speaker labels glued on. The target is the real show. Work one episode per pass, never bulk-complete.

Read first, in this order

1. The skill: ~/.claude/skills/horror-vanguard-reviewer/SKILL.md
2. Three full corpus transcripts from ~/Projects/horror-vanguard-tp/txt/ (target quality, 20 episodes, live at https://drwu-htmls.vercel.app/horror-vanguard-tp). Read S2E1 in full, it is the model.
3. The current output, which is the problem: https://drwu-htmls.vercel.app/horror-vanguard-reviewer-demo and https://drwu-htmls.vercel.app/horror-vanguard-community-s6e12
4. The gate: ~/Projects/horror-vanguard-tp/hv_review_gate.py, plus STYLE_BIBLE.md and style_stats.json (corpus averages ~6,200 words per episode across 124,722 words).

The diagnosis, verified 2026-07-02, keep this frame

Put the real S2E1 transcript next to the generated s6e12 sample and the gap is six specific failures, not a vibe:

1. Templated openers. Every generated review opens with the same "Hello scriptnerd superfans, welcome back" block, nearly word for word. The real show never opens the same way twice. S2E1 opens with the hosts inside the episode's cliffhanger: Ash was "shot three times at point blank range" and insists "it is fine, it is so fine," playing Cooper's situation as a host bit for two full minutes ahead of any welcome. The SKILL.md opener template is the direct cause. Kill it.
2. No bits. The real hosts commit to sustained comic premises pulled from the episode (the slow thumbs up, the stolen ring, the log lady of podcasting) and pay them off with callbacks minutes later. The generated hosts trade polite observations. Nothing recurs, nothing pays off.
3. Essay narration instead of conversation. Generated Ash says "the episode starts with a fake superhero commercial, which is the first smart choice." That is a reviewer describing craft from outside. The real hosts inhabit the episode, talk to each other, derail, and return.
4. No friction. The generated hosts agree with each other in every exchange. A two-hander needs at least one contested read per episode where neither host folds in the next turn.
5. Symmetrical turn shapes. Generated turns are tidy 3-to-5-sentence paragraphs alternating like clockwork. Real turns range from a four-word interjection to a 300-word digression, with fragments, self-interruptions, and abandoned sentences.
6. Half the length. Generated output runs 1,800 to 3,700 words. The corpus averages ~6,200. The compression is where the texture died.

Root cause: the gate measures compliance (banned words present or absent, required moves checked off) and the drafting step writes to the checklist. Checklist-driven drafting produces exactly this: correct, stripped, dead.

What to build

1. Cold-open engine. Replace the fixed opener template in SKILL.md with a rule: every review opens with a different in-world bit derived from that episode's cliffhanger, iconic image, or strangest object. The hosts play the bit straight for 2-4 exchanges ahead of any welcome. Keep only these constants: the audience is "scriptnerd superfans," Ash and John are the hosts, no horror branding. Add a uniqueness check: the first 60 words of a new review must not fuzzy-match the first 60 words of any prior output in ~/Projects/_outputs/horror-vanguard-reviewer/.
2. Bit architecture step. Ahead of drafting, extract 2-3 comic premises from the episode material itself, write them down as a bit sheet (premise, who carries it, where the callback hits). Every bit must recur at least once, at least 800 words after its first appearance.
3. Conversational texture rules. Turn-length distribution must be ragged: at least 15% of turns under 15 words, at least 10% over 150 words. One host finishes or hijacks the other's thought at least twice. At least one digression leaves the episode entirely and returns with a reason.
4. Mandatory disagreement. One contested read per episode. The hosts hold different positions for at least three exchanges. Resolution optional.
5. The discourse move. Name one piece of fan discourse about the episode and dispatch it the way the corpus does ("we will have the owls-aren't-what-they-seem discourse in just a minute").
6. Length floor. 5,000 words minimum, drawn from beat coverage, not padding.
7. Rebuild the gate. Keep the hard bans exactly as they are (branding regex, creeps rename, anti-AI corpus gate, em-dash, ellipsis). Delete the required-moves checklist scoring. Replace it with distribution checks against style_stats.json: turn-length variance within corpus range, question rate, opener uniqueness vs prior outputs, bit-callback presence (a repeated proper-noun or phrase cluster separated by 800+ words), word count floor. The gate should fail a review for being too regular, which is the actual failure mode.
8. Corpus contact per run. The drafting step must quote its 3 sampled transcripts' bit structures into the working notes first, then draft. No drafting from the style bible summary alone.

Keep unchanged

The horror-branding strip table, the scriptnerd superfans rename, the anti-AI voice rules (no em-dash, no ellipsis, no contrastive however/though/yet/while, the 132 tells), DeepSeek-first model routing, the render-and-ship pipeline through render_review_html.py and the lean deployer.

Acceptance test

Regenerate the Community S6E12 review with the rebuilt skill. Ship it beside the old one at /horror-vanguard-community-s6e12-v2 with HTML + md. It passes when: new gate green, 5,000+ words, opener shares no 60-word prefix with any prior output, 2+ bits each with a callback, 1+ contested read running 3+ exchanges, and a human reading the first 500 words cannot tell it apart from corpus texture on turn-shape alone. Update SKILL.md to match what you built, then write the feedback memory so the fix propagates.

One episode per pass. Run the fraud check ahead of any done-claim. Ship everything through ship.py and paste the proof block.