简体中文 | English
Giggle Generation Scripts
Organizes text with "high-density conflict + black humor + subtext-heavy dialogue + narrative reversals". Does not copy specific film scenes or lines.
No Retry on Error: If script execution encounters an error, do not retry. Report the error to the user directly and stop.
Input Collection
Prioritize collecting:
- - Genre and era: Republican China, contemporary, near-future, etc.
- Core conflict: power, money, identity, revenge, misunderstanding
- Protagonist goal: what they want, what they fear losing
- Character relations: allies, opponents, double-dealers
- Target length: short film, single episode, feature
If missing, fill in and state assumptions explicitly.
Input Conflict Handling
Check for conflicting user inputs first. If conflicts exist, handle in this order:
- 1. Tone conflicts (e.g. "light comedy" + "extremely dark tragedy")
- Character conflicts (e.g. "pure good" + "actively evil protagonist")
- Length conflicts (e.g. "5-minute short" + "20 full scenes")
Rules:
- - Prefer the user's latest constraint; if still inconsistent, offer 2 alternative directions, then continue.
- When info is missing, add at most 3 key assumptions and list them in a "Assumptions" section before the body.
- Do not treat unconfirmed settings as given facts.
Output Mode Selection
Choose mode by user goal and context length. Default: "standard".
- - Quick: Synopsis + character bios (for direction-setting)
- Standard: Synopsis + character bios + scene outline + at least 3 full scene scripts
- Long: Synopsis + character bios + scene outline + 6–10 full scene scripts
Output Order
Output strictly in this order:
- 1. Synopsis (300–600 characters)
- Character bios (3–8 people; each: surface identity / true motive / relational tension / speech style)
- Scene outline (8–20 scenes; each: scene-location-time-conflict core-turning point)
- Scene scripts (at least 3 full examples)
Scene Script Serial Output Protocol
When outputting scene scripts, use serial interaction. Do not send all at once:
- 1. Always start from scene 1 (S01).
- Output only one complete scene at a time (dialogue, action, staging, hook).
- At the end of each scene, always ask: "Continue to next scene (S0X)?"
- Output the next scene only after explicit user confirmation.
- If the user requests "output all at once", switch to batch mode after confirming first.
Scene Script Format
Use a uniform template per scene:
CODEBLOCK0
Style Rules
Dialogue Technique (must follow)
- - Characters never answer directly; use反问, analogy, topic shifts
- Discuss life-and-death matters in everyday tone ("tea's cold" = "you're dead")
- Each exchange is a power contest: answering = blocking, deflecting = dodging, counter-question = strike
- In three-person scenes, the third character's lines pace the conflict
- Forbidden: characters stating emotions directly ("I'm angry", "I'm scared")
- Forbidden: explanatory dialogue ("You know, back then that thing was...")
- Forbidden: characters summarizing for the audience ("So you mean...")
Single-Scene Rhythm Formula
Each scene advances in 4 beats:
- 1. 【Probe】Both sides feel each other out with idle talk (1–2 rounds)
- 【Probe deeper】One side suddenly hits the real topic (1 round)
- 【Reversal】The probed side turns it around; power flips (2–3 rounds)
- 【Cliffhanger】Third party or accident interrupts; leave a hook
Every scene must have at least one relational shift: probe→threat, ally→suspicion, or power reversal.
Black Humor Technique
- - Core: serious situation + oddly calm/ordinary phrasing
- Characters worry about trivial things when in grave danger (discussing tea varieties with a knife at the throat)
- Discuss lives in business tone ("Thirty-six lives, wholesale or retail?")
- Absurdity comes from logically consistent nonsense—each line alone sounds "reasonable", together it's absurd
- Forbidden: internet memes, puns, slapstick (not Stephen Chow; Jiang Wen style)
Language Fingerprint Rule
- - Each character's speech pattern must be recognizable in every line
- Cover character names; the lines alone should identify who speaks
- After writing: randomly pick 3 lines—can you tell who said them by tone? If not, rewrite.
Interruption Rhythm
- - Dialogue must be "dense"—responses within a breath, like ping-pong
- Two-person: back-and-forth without pause; three-person: like passing a ball, third can interrupt anytime
- Allow interruption; mark cut-off lines with "—"
- At least one interruption per scene (B starts before A finishes)
- Short lines (5–15 chars) mainly; occasional long line to reset rhythm
- Forbidden: each person delivers a long monologue in turn (that's speechifying, not dialogue)
Information Density (two layers per scene)
- - Each scene's dialogue carries at least two layers: surface topic + real topic
- If a scene advances only one thing, density is low—add a second layer
- Add by: dialogue says A, action/props reveal B; or dialogue literally discusses A while subtext is B
- Self-check: remove all action cues; can the audience feel "they're not just talking about this" from dialogue alone? If not, rewrite.
- Reference: 10 lines per scene should advance at least 2 info points + 1 relationship change
Word Precision
- - After each line: can you cut one character? Cut it.
- Can you replace with a more precise word? "这位大爷" vs "这位爷" are two different people; "走" vs "滚" are two attitudes
- Lines should "taste" when spoken: use punchy words, avoid formal and cliché phrases
- Prefer verbs: specific action over abstract description ("He slaps chopsticks on the table" beats "He's angry")
- Forbidden: literary tone ("岁月如歌"), broadcaster tone ("让我们共同见证"), internet slang ("绝绝子", "yyds")
Other Style Requirements
- - Every 2–3 scenes: one info reversal; prefer action over narration
- Action and staging serve story: position, gaze, noise, props must drive conflict
Quality Checklist
Before output, verify:
- - All four parts present and in order
- Character motives interlock, not isolated
- Each scene has a goal and a change
- Dialogue is speakable and distinguishes voices
- No direct copy of existing film scenes or lines
- Each scene follows 4-beat rhythm (probe→deeper→reversal→hook)
- Black humor present (everyday tone for dangerous matters)
- Can you distinguish characters by tone with names hidden?
- No forbidden items (direct emotion, explanatory dialogue, summarizing for audience)
- Dialogue dense enough; interruptions present (not turn-taking)
- Two layers of information per scene (surface + real)
- Lines trimmed; deleted what can be deleted; words concrete and strong
Revision Loop
When iterating, do at most 2 focused revision rounds; each round only one dimension:
- - Conflict intensity (more restrained / sharper)
- Dialogue tone (more subtle / more pointed)
- Staging (more static / more dynamic)
Each round: first output "this round's changes (up to 3)", then the revised excerpts, not a full rewrite.
Example Reference
To adapt quickly, read references/examples.md and replace setting/characters per the user's topic.
简体中文 | English
咯咯笑生成脚本
以高密度冲突+黑色幽默+潜台词密集对话+叙事反转组织文本。不复制特定电影场景或台词。
出错不重试:如果脚本执行遇到错误,不要重试。直接向用户报告错误并停止。
输入收集
优先收集:
- - 类型与时代:民国、当代、近未来等
- 核心冲突:权力、金钱、身份、复仇、误会
- 主角目标:他们想要什么,他们害怕失去什么
- 人物关系:盟友、对手、两面派
- 目标长度:短片、单集、长片
如果缺失,自行补充并明确说明假设。
输入冲突处理
首先检查用户输入是否存在冲突。如果存在冲突,按此顺序处理:
- 1. 基调冲突(例如轻喜剧+极度黑暗悲剧)
- 人物冲突(例如纯粹善良+主动作恶的主角)
- 长度冲突(例如5分钟短片+20个完整场景)
规则:
- - 优先采用用户的最新约束;如果仍不一致,提供2个替代方向,然后继续。
- 当信息缺失时,最多添加3个关键假设,并在正文前的假设部分列出。
- 不要将未经确认的设置视为既定事实。
输出模式选择
根据用户目标和上下文长度选择模式。默认:标准。
- - 快速:梗概 + 人物简介(用于方向确定)
- 标准:梗概 + 人物简介 + 场景大纲 + 至少3个完整场景脚本
- 长篇:梗概 + 人物简介 + 场景大纲 + 6-10个完整场景脚本
输出顺序
严格按此顺序输出:
- 1. 梗概(300-600字)
- 人物简介(3-8人;每人:表面身份/真实动机/关系张力/说话风格)
- 场景大纲(8-20个场景;每个:场景-地点-时间-冲突核心-转折点)
- 场景脚本(至少3个完整示例)
场景脚本序列输出协议
输出场景脚本时,使用序列交互。不要一次性全部发送:
- 1. 始终从场景1(S01)开始。
- 每次只输出一个完整场景(对话、动作、调度、钩子)。
- 在每个场景结尾,始终询问:继续下一个场景(S0X)?
- 仅在用户明确确认后输出下一个场景。
- 如果用户要求一次性全部输出,在确认后切换到批量模式。
场景脚本格式
每个场景使用统一模板:
text
【场景】S03
【地点/时间】县衙大院/夜晚
【人物】马走日,黄四郎,账房先生
【场景目标】马走日想套出银票去向;黄四郎想反设局。
【动作与调度】
- - 马走日围着石桌走了半圈,始终不坐。
- 黄四郎背光而立;账房先生略退两人身后。
- 远处鞭炮声在提到规矩时打断。
【对话】
马走日:这院子里的风,刮得跟刀子似的。
黄四郎:风不狠,人站不住。
账房先生:二位爷,茶凉了。
马走日:茶能热,账本不能热——热了就得见血。
【钩子】
账房先生袖口滑出半截银票存根,马走日看见却装作没看见。
风格规则
对话技巧(必须遵循)
- - 角色从不直接回答;使用反问、类比、转移话题
- 用日常语气谈论生死攸关之事(茶凉了= 你死定了)
- 每次交锋都是一场权力较量:回答=阻挡,回避=躲闪,反问=反击
- 三人场景中,第三人的台词为冲突定节奏
- 禁止:角色直接表达情绪(我很生气、我很害怕)
- 禁止:解释性对话(你知道,当年那件事是……)
- 禁止:角色替观众总结(所以你的意思是……)
单场景节奏公式
每个场景按4个节拍推进:
- 1. 【试探】双方用闲话互相试探(1-2轮)
- 【深入】一方突然切入正题(1轮)
- 【反转】被试探方扭转局面;权力翻转(2-3轮)
- 【悬念】第三方或意外打断;留下钩子
每个场景必须至少有一个关系变化:试探→威胁,盟友→猜疑,或权力反转。
黑色幽默技巧
- - 核心:严肃情境 + 异常冷静/普通的措辞
- 角色在极度危险时担心琐事(刀架在脖子上讨论茶叶品种)
- 用生意口吻谈论人命(三十六条命,批发还是零售?)
- 荒诞来自逻辑一致的胡言乱语——每句单独听合理,合起来荒诞
- 禁止:网络梗、双关语、闹剧(不是周星驰风格,是姜文风格)
语言指纹规则
- - 每个角色的说话方式必须在每句台词中可识别
- 遮住角色名;仅凭台词应能分辨是谁在说话
- 写完后:随机挑3句台词——能凭语气判断是谁说的吗?不能则重写。
打断节奏
- - 对话必须密集——一息之间回应,如乒乓球对打
- 两人:来回不停顿;三人:如传球,第三人可随时打断
- 允许打断;用—标记被打断的台词
- 每个场景至少一次打断(B在A说完之前开始)
- 主要用短句(5-15字);偶尔用长句重置节奏
- 禁止:每人轮流长篇独白(那是演讲,不是对话)
信息密度(每个场景两层)
- - 每个场景的对话至少承载两层:表面话题 + 真实话题
- 如果一个场景只推进一件事,密度低——添加第二层
- 添加方式:对话说A,动作/道具揭示B;或对话字面讨论A,潜台词是B
- 自查:去掉所有动作提示;仅凭对话,观众能感觉到他们不只是说这个吗?不能则重写。
- 参考:每个场景10句台词应推进至少2个信息点 + 1个关系变化
用词精准
- - 每句台词后:能删一个字吗?删掉。
- 能用更精准的词替换吗?这位大爷 vs 这位爷是两个人;走 vs 滚是两种态度
- 台词说出来要有味儿:用有冲击力的词,避免正式和陈词滥调
- 偏好动词:具体动作优于抽象描述(他把筷子拍在桌上胜过他很生气)
- 禁止:文学腔(岁月如歌)、播音腔(让我们共同见证)、网络用语(绝绝子、yyds)
其他风格要求
- - 每2-3个场景:一次信息反转;偏好动作而非叙述
- 动作与调度服务于故事:位置、目光、噪音、道具必须推动冲突
质量检查清单
输出前,验证:
- - 所有四部分齐全且顺序正确
- 角色动机环环相扣,非孤立
- 每个场景有目标和变化
- 对话可口语化且区分声线
- 没有直接复制现有电影场景或台词
- 每个场景遵循4拍节奏(试探→深入→反转→悬念)
- 黑色幽默存在(用日常语气说危险之事)
- 遮住名字后能凭语气区分角色吗?
- 没有禁止项(直接情绪、解释性对话、替观众总结)
- 对话足够密集;有打断(非轮流发言)
- 每个场景两层信息(表面+真实)
- 台词精简;删掉能删的;用词具体有力
修改循环
迭代时,最多进行2轮聚焦修改;每轮只针对一个维度:
- - 冲突强度(更克制/更尖锐)
- 对话语气(更含蓄/更犀利)
- 调度(更静态/更动态)
每轮:先输出本轮修改(最多3项),然后输出修改后的节选,而非全文重写。
示例参考
为快速适配,阅读references/examples.md,并根据用户主题替换设定/角色。