Wider ideas, scores that follow the film, recipes when you want them

[ZLHad]

Step 2 of the plan; step 1 is #62 (merged as c09efb8). This branch was stacked on #62; main is now merged in, and the diff against main is this step only.

Why

Step 2 of the maintainer-approved plan: give the model room to think wide instead of steering it to one house look, without loosening the floors.

• The concept step had no push away from the median. It asked for 5–8 ideas, but nothing checked whether they were the ideas everyone has, and the median shows in the output: four of the five showcase films use a yellow or orange accent, and four sit on a dark ground.
• The scoring layer measured every film against short-form motion design: "前 2 s 有没有让人停下来", "每 2–4 s 有没有新东西", "像不像专业动效". A slow explainer, a letter or a hand-drawn piece loses points for being what it is. studio's "8 项都 ≥ 8" was never met: the intro film's five rounds peaked at 8 7 8 7 7 7 9 8 (round 4), and the user ended the loop after round 5 (gates ②③ had been waived and reviewed by an independent agent). A lighter read path: CLAUDE.md as an index, one home for taste defaults, the audio reference split out #62 already made ≥ 8 studio-only.
• Recipes were the start of every shot ("分镜时先查 recipes", "每个镜头写配方 id 或'自创：理由'"), the same shape the style presets had before Style presets are references, not templates #48.

What changes

Divergence (playbook/12-ideation.md §3, §4, §7; CLAUDE.md step 4; playbook/quick.md; playbook/01's effort table; templates/NOTES.md)

• Four questions before the ideas: what the subject itself looks like, what the emotion looks like as one frame, where and how it is watched, and what every other video on this subject looks like (that is the anti-pattern).
• A rough probability per idea ("how likely is another model to think of this first"), written with the idea ("一句话 · 约 0.3"), kept out of the review page; fewer than two under one in ten means think again. Two ideas whose rough sketches match count as one. When the user already gave the concept, ideas and probabilities are skipped; the four questions still set style and hook.
• These three steps are translated and rewritten from HyperFrames' pitch round (skills/hyperframes/references/pitch-round.md, commit a46095f; Apache-2.0, Copyright 2026 HeyGen, Inc.). The section says so, with the licence copy (recipes/LICENSES/Apache-2.0-hyperframes.txt, already in the repo); ACKNOWLEDGMENTS lists it under licence scope. HyperFrames has no NOTICE file at that commit. The emotion and playback examples are our own. The probability step is also what Verbalized Sampling tested (Zhang et al., ICML 2026, arXiv:2510.01171: asking for candidates with probabilities raises creative-writing diversity 1.6–2.1× over direct prompting; verified on the arXiv abstract).
• At least one concept card comes from the long tail (the recommendation still goes by §5's tests, not novelty); cards are written with equal care so the recommendation doesn't anchor; "the framing of card 2 with the opening of card 1" is a normal answer.
• quick answers the four questions too, and writes down the most typical approach it skipped; the delivery lists that line. "Nobody is watching" runs are where this matters most.
• NOTES.md gets a short 点子 section for the draft (the desk reader parses sections by title, unaffected).

Scoring (templates/TASTE_CHECKLIST.md, playbook/01 and 02, both READMEs, type 09, engines/editing.md)

• The reviewer is "a harsh director" (not "motion director"), and gets the BRIEF's Watch on and Platform lines, STYLE.md's Motion tokens and the opening 2 s strip (plus the first 10 s for a slow opening), along with the concept line, Pace and taste overrides.
• Motion is scored against the motion the film's STYLE.md declares (smooth long tails, mechanical steps, cartoon overshoot, handheld shake…; playbook/03's defaults when it declares none), not "does it look like professional motion design". Variety uses Design notes: v0.3 roadmap #18's yardstick (3 s short, 8 s explainer), or the meaning of each hold in a film meant to be slow. Hook asks what fits where it is watched: the first 2 s in a feed; for an explainer, a letter or an MV someone chose to open, whether the first seconds set up a question or an image.
• Passing per level: standard has no score line (one round, the worst three fixed, scores and leftovers to gate ③). studio needs the five craft scores (target-screen legibility, motion, composition and polish, accuracy, sync; sync is N/A for a silent film) at 7 or above; concept, hook and variety are judged against the film's own concept with no line, and the lowest goes to gate ③ with the reviewer's fix. An authorised unattended run has no gate ③, so those go into the delivery notes; an issue capped at 5 fixes doesn't keep the loop going. Floor FAILs block delivery at any level, as before. Caps (10 rounds, 5 per issue) unchanged.
• The scoring layer's source note records the change, why (the v5 numbers above) and that the maintainer agreed to it.

Recipes (recipes/README.md, templates/STORYBOARD.md, CLAUDE.md, type 03): a reference for shots you're unsure of. The recipe column stays empty for your own shots (no "自创：理由"); skeletons are optional at standard; studio's A/B at hook and climax may compare any two approaches. bin/vh recipes check passes.

Precedence (CLAUDE.md "规则冲突时", now a numbered list; mirrored in recipes/README and playbook/12 §6): user (only four floors can be waived by the user's own words: text size, reading time, silence and the gates, recorded verbatim and dated in REVIEW.md; facts, flashes, consent, platform and legal lines, third-party licences and keys are not waived by a remark) > floors: hard rules, the floor list, the human gates and who may lower effort, and any rule in any doc on facts, licences, legibility, privacy and consent, platform and legal lines > the concept card the human approved at gate ①, and the project's STYLE.md, where they differ from the defaults only with the reason written down first in DECISIONS.md (added after gate ①: judged by the reviewer) > the type doc and its engine guide > recipes > playbook > references.

Type docs: type 02 calls "Kurzgesagt meets Fireship" one of the common ways to make this kind of film 【综合】: a start when there is no concept, and one answer to "what does every other video look like" when there is. Type 03 says the same of its two registers, Linear-style being the community's most common look 【综合】. playbook/03 §0's list doubles as part of that answer. playbook/01's look-dev may vary the style or the concept itself when the note is "it feels wrong", not only effect strength.

Not changed

Floors, gates, review caps, bin/vh code. No style presets were added (the warm-accent skew in styles/ is noted for future additions). Step 3 (rule hygiene: 【单例】 tags, a read budget printed by CI) is later.

Verification

• tools/ci.sh --committed with shellcheck and pyflakes on PATH: all checks passed.
• VH_BASH=/bin/bash tools/ci.sh --committed: all checks passed.
• git grep for leftover "≥ 8" score lines, "动效导演", "先套骨架", "自创：理由" outside history (CHANGELOG, showcase, research notes): none left except the scoring layer's own source note and styles/README.md's account of how the swatches were reviewed (history, kept).
• Needs Mac verification: none (docs only).

Review

An independent reviewer (fresh context) asked for changes: the floor layer of the new precedence order was too narrow (consent, privacy and platform rules outside type docs, and the human gates, could have been outranked by STYLE.md), the reviewer lacked the inputs the new hook / variety / motion criteria need, the intro film's history was wrong, unattended runs had nowhere to put "to gate ③" items, the pitch-round text was partly a translation rather than "our own words" and lacked an Apache-2.0 note, and one type-03 sentence didn't read. All fixed in e7329d9, along with most of its suggestions. The same reviewer re-checked 4176899 and approved; its two pre-merge suggestions (the scope of a user's waiver, CHANGELOG and PR wording) and three small ones are in 204ca8d, which it also approved; a09613c applies its two optional one-line suggestions (the waivable floors are a closed list of four, facts are protected, a waiver is dated).

Merging

#62 squash-merged as c09efb8 and GitHub retargeted this PR to main. main was merged into this branch (4176899, no history rewrite): main's only new commit since the fork point is #62, whose tree is byte-identical to #62's approved head b976e9d, already merged here, so the conflicts were only lines this PR rewrote on top of #62. Taking this branch's side leaves the tree byte-identical to the pre-merge head (git diff d5b0efb 4176899 is empty).