script-judges

# Script judges — tear the lines apart before you spend a cent Video models render a bad script exactly as pretty as a good one. Once the footage passes as human, retention lives or dies on whether the LINES sound spoken or written. So the script gets torn apart first — by narrow, bad-tempered specialists, not one polite generalist. **How to use**: give the panel a script (numbered lines or per-shot voiceover). Run ALL four judges, each strictly from its own angle. Then rewrite. Then run the panel again. A script only graduates when every judge passes it. ## The four judges Play each judge as a separate, single-minded reviewer. Harsh beats kind; a judge that finds nothing is suspicious. ### 1. 节奏官 — the pacing judge Owns exactly one thing: does the viewer stay? - The first line is a death-penalty checkpoint: if it wouldn't stop a thumb mid-scroll, the script is dead — no other virtue can save it. - Every line must make the next line wanted. Flag any sentence that answers a question nobody asked. - Flag long winding sentences and low-information filler. Spoken lines are short. ### 2. 口语官 — the spoken-voice judge Owns exactly one thing: does it sound SAID, not WRITTEN? - Hooks start mid-conversation — the viewer scrolled in halfway. "大家好 / 今天给大家介绍 / Hi everyone" openers are instant kills. - Filler words ("就是", "说真的", "I mean") and one half-finished thought are GOOD — sparingly, once or twice per script. - Written-prose connectives ("因此 / 综上所述 / therefore / in conclusion") are kills. Punchline endings and life-lesson closers are kills. - CTAs must sound offhand ("反正链接我放这了" / "anyway, link's down there") — never a slogan. - The test: read it aloud. Anything that sounds like copy gets rewritten as speech. ### 3. 创意官 — the freshness judge Owns exactly one thing: has the viewer seen this before? - Flag template openers, worn-out memes, and any line that every competing video in the niche would also say. - Flag "safe" claims that could describe any product. Specifics beat adjectives; one concrete detail beats three superlatives. ### 4. 结构官 — the structure judge Owns exactly one thing: does the whole thing build and land? - Setup → escalation → landing must actually escalate. Flag scripts that plateau after the hook. - Endings that land on a lesson, a moral, or a golden line are violations — end where the thought trails off or gets interrupted. - A slogan-shaped CTA is a violation (that's also the voice judge's kill — they agree on this one). ## Feeding the judges (calibration corpus) Judges are only as sharp as what they're fed. Before serious use, paste calibration material below each judge's section in your working copy: - Transcripts of creators whose videos actually stop YOUR thumb — hooks, full voiceovers, topic choices. - A "proven scripts" baseline: scripts (yours or studied) behind videos with real performance numbers. **The iron rule — easy to get backwards**: at least one judge (the spoken-voice judge) must be fed ONLY real creators' raw transcripts — never award-winning ad copy. UGC lines are not supposed to sound well-written; a judge calibrated on polished copywriting will "improve" every script into an ad, which is the exact failure this panel exists to prevent. ## Rewrite rules - Keep the meaning, the selling points, and every factual claim. Change only HOW it's said. Never add new efficacy/price claims (ad-law red line). - Keep each rewritten line within ±20% of the original length — voiceover duration is usually pinned to shot slots by TTS. - Rough beats polished. If it reads like copy, rewrite it as speech. - Across a batch, vary the person, the setting and the phrasing patterns — repetition is the biggest tell. - If the hook wouldn't hold you, say so BEFORE anything gets generated. If a claim isn't supported by the product, flag it. ## With ClipForge ClipForge ships this panel built in: the script page's ⚖️ judge-panel button runs all four judges in one LLM call and applies rewrites through `PATCH /api/project/{id}/scripts` (`shotTexts`). As an agent you can do the same: `POST /api/project/{id}/script-judge { scriptId, llmConfig }` returns `{ verdicts, rewrites, summary }`, then PATCH the rewrites you accept. Judging is cheap (one text call); generation is not — always judge first.