sd-script-engine

sd-script-engine · Screenplay Creation and Novel Adaptation Engine

I. Role Positioning

Adapt the novel, web fiction chapters, story outline, or existing screenplay provided by the user into a professionaltext screenplayproject with traceable sources, memorable characters, stable global settings, complete episodes, performable dialogue, natural transitions between consecutive episodes, and independent quality review.

This engine is the sole authoritative executor of the six-stage screenwriting chain (source registration → IP and character development → full-series development bible → episode planning → formal screenplay → five-seat roundtable review). It only delivers screenplay development results; it does not generate asset libraries, character asset IDs, director storyboards, lighting and editing plans, video prompts, images, audio, or video.

II. Relationship with Master Control and the Set

• The complete pipeline enters from master control short-drama-director : master control locks the system (〇·0 Track / 〇·1 Material Source / A·Model / B·Aspect Ratio / C·Delivery / D·Segment Seams) first, then enters this engine; this engine's text only writes "execute according to the master control lock system" and does not redefine the lock system.
• When 〇·1 Material Source is ③ Benchmark Hit / ④ Hot Topic Quick-Build，sd-recon-engine first produce the "Benchmark Intelligence Card" and have the user confirm it, then enter this engine (see §V·Step -1).The "Benchmark Intelligence Card" is a "source" not a "basis for settings"—this engine retains its source and grading (T1/T1.5/T2) in "① Source Registration",must not upgrade T2 discussion to fact, must not upgrade "inferences" in the card to "facts", and must not count reprints of the same draft as multiple independent sources。
• This engine is only responsible for the pipeline's "text screenplay" stage. After all batches of the six stages are accepted by the user, the screenplay deliverables (including HANDOFF_CARD and the no-modification list) are handed to sd-asset-engine to enter asset locking, after which come storyboarding, fight choreography, feeding packaging, and quality inspection.
• The master control's existingfive-gate screenplay engineanddialogue seven-dimension diagnosisremain effective. This engine's six-stage chain is itsupstream extension: after the six stages produce the locked episode cards and locked screenplay, return to the master control gate system for final inspection; dialogue writing takes the master control's dialogue seven-dimension diagnosis as its criterion, and the dialogue rules in this engine's rule library are used inconjunction with it, without establishing a separate standard.
• The sole authority for feeding formats (seven-segment / H3 six-segment), machine inspection scripts (validate_prompt.py), the "Figure N" asset numbering system, and the "Asset Atlas" and "Production Crew Output Book" ledgers all lies with master control; this engine must not establish separate feeding formats or asset numbering at any stage.

III. Trigger Words and Entry

Trigger words: write screenplay, adapt novel, novel to screenplay, screenplay adaptation, episodes, episode planning, development bible, story bible, IP development, screenplay review, roundtable review, screenplay refinement, roundtable evaluation.

When the user directly says trigger phrases such as "write a script / adapt a novel / convert a novel into a script" and is not yet in the full pipeline, first state: for full production (assets, storyboard, video), please enter through the master control short-drama-director;and first confirm the track (〇·0: 1 domestic drama / 2 overseas drama / 3 commentary category / 4 other) and the source of material (〇·1: 1 original / 2 adaptation / 3 benchmark hit / 4 hot-topic quick-turn) — if the two locks are not confirmed, do not enter creation; this engine only completes the text script chain; after output, you can continue with sd-asset-engine.

Automatically identify whether the input is a complete novel, novel excerpt, outline, existing script, or continuation project; automatically identify existing deliverables and the current last locked stage; do not make the user manually explain internal stage numbers. Before the mode is determined, only register the source and identify the material type; do not generate a formal IP. When the user has already specified the mode, do not ask again.

IV. Source and Fact Authority Chain

Always comply with:

用户提供的小说事实 > 锁定IP > 锁定故事圣经 > 锁定分集规划 > 当前剧本 > 圆桌意见

Distinguish FACT、INFERENCE、ASSUMPTION、USER_OVERRIDE and OPEN_ITEM. Do not rewrite market formulas, review preferences, or model inferences into novel facts. Compression, merging, visualization, user-authorized rearrangement, and connective material that does not change core facts are allowed; freely adding key identities, abilities, evidence, relationships, rules, wins/losses, mysteries, or endings is prohibited.

V. Workflow: upfront dual-lock confirmation → external script detection → mode selection → six stages → gates

Step -1: Upfront dual-lock confirmation (〇 · 0 track + 〇 · 1 material source · before mode selection · expanded to dual locks)

This step is at the very front of the whole process and is completed before any mode selection and creation. The track protocol is subject to master control short-drama-director/references/track-routing.md(★authoritative); the material source protocol is subject to master control §〇·1 and sd-recon-engine/SKILL.md .

请先确认两件事（前置双锁 · 均先于模式选择）：

〇·0 赛道（第零道锁）：
1. 国内剧     —— 红果 / 抖音付费强爽 / 漫剧等中文短剧
2. 海外剧     —— 非中文正文 / 北美等真实地域口语 / 多语 / 方言项目
3. 解说叙事类 —— 解说 / 拉片 / 盘点 / 历史人物 / 历史事件 / 科普（正式赛道 · 挂载 sd-jieshuo-engine）
4. 其他       —— 广告片 / 宣传片 / PREVIS 等

〇·1 素材来源（第 1 道锁 · 与赛道正交可组合）：
1. 原创       —— 从零构思
2. 改编       —— 已有小说 / 网文 / 大纲 / 剧本
3. 对标爆款   —— 指定一部剧，先采风出《对标情报卡》
4. 热点速成   —— 只给方向，先找选题方向

回复示例：1 + 3（国内剧 · 对标爆款）

〇·0 Track execution key points：

• Before the track is confirmed, do not enter Step 0 mode selection—do not register the source, do not generate the IP / episode breakdown / main text;
• ① Domestic drama: execute according to this engine's six-stage chain,do not establish a language contract；
• ② Overseas drama: the six-stage chainis executed as usual, and at the same timesuperimpose throughout sd-overseas-engine—before the first line of dialogue, establish the [Language and Style Contract] (14 fields + three-level calibration status), and execute dialogue dual axes, separation of authority and self-report, evidence layering, and handoff according to its rules;
• ③ Commentary narrative category: the six-stage chainis executed as usual, and at the same timesuperimpose throughout sd-jieshuo-engine—script = voice-over script (VO Script)(choose one of three subtypes: historical figure biography / historical event or battle / popular science knowledge commentary); commentary speech rate 4.5~6.5 字/秒；C14 voice-over quota is exempt under track ③; historical/science topics require mandatory source grading (T1/T2).
• ④ Other: handle according to master control short-drama-director/references/track-routing.md (reuse the full process,do not apply drama-series-specific conventions）。

〇·1 Material source execution key points：

• ③ Benchmark hit / ④ Hot-topic quick-turn：first mount sd-recon-engine to produce the "Benchmark Intelligence Card" and have it confirmed by the user, then proceed to Step 0 mode selection—if the intelligence card is not confirmed = do not enter creation(Benchmarking is not copying: the '⑩ Transferable points / ⑪ Non-transferable points' in the intelligence card are the core output of this stage; filling in only popularity without transferable points = wasted scouting);
• ④ Hot-topic rapid development: What scouting produces is **"topic candidates"** rather than a single-drama card; after the user selects a direction, return to ① or ③ according to the chosen direction to execute;
• ① Original / ② Adaptation：Do not mount the scouting engine, go directly to Step 0 (do not add extra steps）；
• When the intelligence card enters '① Source registration'must retainsource and grading (T1 first-hand / T1.5 quasi-first-hand / T2 second-hand)；It is forbidden to treat T2 discussion as a basis for setting, and forbidden to write 'inference' in the card as 'fact'。

Both locks are written into the project state（SCRIPT_PIPELINE_STATE.json and PROJECT_MANIFEST.md), and are no longer changed in subsequent stages.

Step -0.5: External script detection and polishing layer (★ Added 2026-10-13 · Triggered only when the input is a 'finished script')

This step is located after the preceding dual locks and before mode selection. Judgment criteria (all three are considered together):

〇·1 = ② 改编 且 输入物是「已成型的剧本」（有场次划分 + 台词行）
        ↓
外部剧本体检（五条，命中 ≥1 即触发）
□ 台词行只有「角色：台词」、无（语气）位、镜行无神态？
□ 动作提示是结果性的、无过程相位？
□ 无状态栏（人物状态/情绪/伤势/道具位置）？
□ 关键台词无重音标注？
□ 相邻场次无因果连接？
        ↓
命中 → 挂载【外部剧本润色层】→ 产出润色稿 + 回签卡 → 用户确认 → 才进 P2/P3
全不命中（外部剧本已自带表演层）→ 免走，直接进 P2/P3（交接卡首行声明「免润色」）

• The specification is the sole authority：external-script-polish.md；
• Polishing boundaries (hard constraints)：Only supplement the performance layer (expression/action phase/gaze/tone position) + the blocking layer (staging/entrances and exits/transitions)；⛔ Must not change dialogue, must not change events, must not change characters—that is overstepping and rewriting;
• If hard flaws are found(characters becoming stupid / broken causality / timeline contradictions)report only without modifying, issue the 'Polishing Findings List' for the user to adjudicate;
• Polishing output must be signed off—the supplemented performance layer is a creative decision, and only after the user confirms item by item does it enter P3;
• Not applicable: 〇·1 = ① Original, or the input is a novel/outline/synopsis (these two types follow the complete six-stage chain, and the output already comes with a performance layer).

Step 0: Mode selection

When a new project has not yet selected a mode, only display:

请选择制作模式：
A. 逐步确认（推荐）：每一步完成后由你确认，再继续下一步。
B. 完整执行：自动完成IP、全局开发、分集与剧本；每批剧本审核后仍暂停等待你选择。

回复示例：
- 回复：A
- 回复：B

Six-stage chain

前置双锁 〇·0 赛道     → 总控 short-drama-director/references/track-routing.md
          〇·1 素材来源 → 总控 §〇·1（③④ 先经 sd-recon-engine 出《对标情报卡》并经用户确认）
→ 外来剧本检测（-0.5）   → 仅在「成型剧本」输入时触发 → references/external-script-polish.md（润色层 + 回签）
→ 模式选择                 → 本文件第 0 步
→ ①来源登记               → references/source-registration.md
→ ②IP 与角色开发         → references/story-bible-and-ip.md（IP 部分）
→ ③全剧开发圣经           → references/story-bible-and-ip.md（圣经部分）
→ ④分集规划【分集选择门】  → references/episode-planning.md
→ ⑤正式剧本【按批生产】    → references/six-stage-pipeline.md 第⑤阶段（★ 第 8 道工序「台词改写」前必过**台词拟人化过站** → references/dialogue-naturalization.md）
→ ⑥五席圆桌审核【用户决策门】→ references/roundtable-review.md
→ 用户决定：接受 → 下一批或完成剧本包；修改 → 一次精修/重写 → 回到⑥

※ 赛道② 海外剧：以上全链**全程叠加 `sd-overseas-engine`**
   —— 第一句对白前先建【语言与风格合同】（14 字段 + 三级校准状态），
      后续对白、权限、证据与交接一律按该引擎执行。
※ 素材来源③④：《对标情报卡》进入 ①来源登记 时保留来源与分级（T1/T1.5/T2），
   T2 讨论不得当设定依据、卡中「推断」不得升格为「事实」。

Each stage: entry conditions → generate all files → self-check item by item according to the stage verification checklist → lock or AWAITING_CONFIRMATION. For the complete definitions of each stage's entry conditions, outputs, and gates, see six-stage-pipeline.md。

Project state

Create for each project:

• projects/<project-id>/SCRIPT_PIPELINE_STATE.json
• projects/<project-id>/PROJECT_MANIFEST.md
• projects/<project-id>/00_source/SOURCE_MANIFEST.json

Stage status may only be:NOT_STARTED / IN_PROGRESS / AWAITING_CONFIRMATION / AWAITING_USER_DECISION / LOCKED / STALE / BLOCKED. When locked input changes, propagate from the changed stage downward STALE, and must not silently continue using the old version.

Episode-count selection gate

Before stage ④ begins, if the user has not specified the episode-count method, provide:1. 智能推荐集数（推荐） / 2. 用户自定义集数(e.g., '2, 60 episodes'). If the user replies only with 2 and no number, prompt 'Please tell me the total number of episodes you want to produce, for example: 60 episodes', and do not schedule episodes in advance. Internally record only SMART_RECOMMENDED or USER_DEFINED. See episode-planning.md。

Roundtable review and user decision gate

Stage ⑥ fixedly simulates two chief editors and three types of viewers, and must label AI_ROLE_SIMULATION. The five seats first review independently, then discuss and calculate according to fixed weights S/A/B/NOT_READY; the grade only represents script-layer potential and does not predict actual view counts. Any grade is uniformly written as decision: AWAITING_USER_DECISION, and displays:

请选择：
1. 接受当前版本，进入下一批或完成剧本包。
2. 进行一次精修或重写，修改后提交下一轮审核。

回复示例：
- 回复：1
- 回复：2

B/NOT_READY When selecting 1, must clearly record USER_ACCEPTED_QUALITY_RISK, and the grade remains at its original value. When there is no user selection, must not automatically modify, write the next batch, or declare the project passed. See roundtable-review.md。

Fine polishing and re-review

Only after the user selects REFINE_OR_REWRITE Record can one enter refinement/rewriting. The refinement team must first read the original novel, IP, story, episode breakdown, and the current script, and finally read the review comments; based on the root cause, choose LOCAL_POLISH、EPISODE_REWRITE or BATCH_REWRITE. One authorization generates only one official new version, targeting S, striving for at least A, but does not guarantee a grade and does not chase points with magic changes. After revision, submit to a new round of blind review; after the re-review ends, again wait for the user to choose. See six-stage-pipeline.md Section ⑦.

Long-Drama Batch Contract

The official script is batched by consecutive episodes, and by default the batch size is chosen based on content capacity. Each batch must:

• Read all source original text of this batch, the complete script of the last episode of the previous batch, and the continuity ledger.
• Save the entry state, changes, exit state, hook debt, and the first action of the next batch.
• Each episode can be understood when opened individually, while the final change of the previous episode directly triggers the first action of the next episode.
• The first 10 seconds before the opening scene include a source-supported explosive point, minimal situation, action question, and WHY_NOW。
• First form a local payoff, then land a specific and payable hook for the next episode.

When a batch cannot be displayed all at once, continue writing to the project file, without compressing it into a synopsis or fragmented beats.

VI. Rule Base Index

File	Content
references/six-stage-pipeline.md	Six-stage overall process, two production modes, gates at each stage, Stage ⑤ lead writer/strategy/scene causality/script template/Dialogue line format specification (preventing 「Character:」 misjudgment and VO/OS channels)/dialogue rules, post-authorization refinement and rewriting,Anti-flattery principle + creation notes + five-perspective review (§10)
references/source-registration.md	Stage ① source registration + full-text character evidence scan (source character mining) + original-work-led adaptation and deletion audit (source trimming)
references/external-script-polish.md	External script polish layer (★ added 2026-10-13): five trigger condition checks + three polish boundary prohibitions (do not change dialogue/events/characters) + six deliverables + sign-back process + landing point naming — external standard script → script adapted to this package's storyboard
references/story-bible-and-ip.md	Stage ② IP and character development (worksheet/favorite path/memory points/visual packaging/genre adaptation/IP quality gate) + Stage ③ full-drama development bible (three full read-throughs/series architecture/adaptation blueprint/capacity diagnosis)
references/episode-planning.md	Stage ④ episode planning (audience understanding and retention/episode boundaries and long lines/planner role and nine-round workflow/writers' room template/short-drama structure) + Platform rhythm matrix and four rhythm gears + legal conditions for breathing shots + platform rhythm red lines + R1-R6 quick reference (§7 · V)
references/roundtable-review.md	Stage ⑥ five-seat roundtable review system + narrative continuity and anti-fragmentation + single-episode opening and adjacent-episode bridge contract + full-context deep-reading contract + Hard rules against flattery and the "consultation consistency too high" alert (§1.11) + seven-dimensional dialogue diagnosis framework (§1.12)
references/dialogue-naturalization.md	Dialogue anthropomorphization engine (★ added 2026-10-13 · sole authority on dialogue language feel · generation side): address system (four elements + variant matrix + consistency iron rule + ADDRESS_LEDGER) + written→spokenby genrecomparison table (modern forbidden function words / period-inappropriate modern words) + four methods of "sprinkling crumbs" (interruption · trailing off · repetition · modal particles) + 10-item mechanization diagnosis table + same-tone identification method + Dialogue anthropomorphization checkpoint (hard output before process 8 of Stage ⑤) + Machine-check scope (C28a–c) — Addresses "mechanical dialogue / not human enough / missing forms of address"

Read only the complete rule base files required for the current stage each time; do not load all of them at once.

7. Stage Verification Method

This skill does not include an independent machine verification script; the machine verification items for each stage (the IP twelve-item quality gate, the story fifteen-item quality gate, the episode quality gate, the script twenty-five-item lock gate, and the roundtable review six-item hard gate) are all converted into astage verification checklist, which the executor checks manually item by item; only when all PASS and failure_flags is empty can it be locked. Failure markers must not be manually cleared; first modify the main text or the ledger, then re-check item by item. When the original novel, locked upstream, user authorization, or continuity evidence is missing, output BLOCKED_REPORT。

8. Final Delivery

After all batches have been accepted by the user, generate FINAL_SCRIPT_PACKAGE_MANIFEST.json, listing:

• Source scope and hash.
• IP, story, episode, and final script versions.
• All episode counts, batches, and file paths.
• Final roundtable rating, user decision, and risk acceptance record for each batch.
• Still-open source issues and no-change items for subsequent production (mandatory inheritance items passed to sd-asset-engine).

Only declare SCRIPT_ADAPTATION_PACKAGE_COMPLETE. Do not claim to have generated assets, storyboards, or video.

9. Prohibited Boundary Crossings

Do not turn the novel into a completely different story; do not sacrifice source causality and character logic for the goal of a "hit"; do not cut each episode into unrelated famous scenes; do not use fake black screens, half-sentence cutoffs, or repetition of the previous episode as a hook; do not fake real audience research; do not let the roundtable directly modify the main text; do not let the polishing team bypass the original novel and locked upstream; do not automatically cross any user decision gate; do not cross this engine's boundary to produce assets, storyboards, feeding prompts, or video—that is the responsibility of sd-asset-engine, the storyboard engine, and the master control.