workflow-creation

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Workflow Creation

工作流创建

Purpose

目的

Design one reliable video-production workflow on the PostPlus workflow platform for UGC-style multi-clip video generation. The user-facing deliverable is a named workflow on the user's account with a saved definition version, ready to quote and launch. Do not produce standalone definition files, exported JSON graphs, audit packets, or post-production notes unless the user explicitly asks.

Quality checks are internal; do not ask the user whether to run QA. Ask only when missing brand, product, reference, workflow-goal, naming, or claim-boundary facts would change the workflow.

在PostPlus工作流平台上设计一个可靠的视频制作工作流，用于UGC风格的多片段视频生成。面向用户的交付成果是用户账户中一个带有已保存定义版本的命名工作流，可直接进行报价和启动。除非用户明确要求，否则不要生成独立的定义文件、导出的JSON图表、审核数据包或后期制作说明。

质量检查为内部操作；无需询问用户是否运行QA。仅当缺少品牌、产品、参考资料、工作流目标、命名或声明边界等信息会改变工作流时，才向用户询问。

Platform Tool Surface

平台工具界面

Workflow authoring is ONE server-validated capability with two front doors. Use whichever the current session exposes — never mix them, and never bypass server validation:

Workspace assistant typed tools:
```
workflow_read
```
(resolve by name via
```
action=list
```
, inspect nodes/slots/versions via
```
get
```
, read runs via
```
runs_get
```
),
```
workflow_author
```
(
```
create
```
a blank or template-seeded workflow;
```
propose_edit
```
structural edits -
```
add_node
```
/
```
update_node
```
/
```
remove_node
```
/
```
connect_nodes
```
- returning an honest validation verdict without persisting),
```
apply_workflow_edit_to_canvas
```
(same operations when the user's live canvas is attached),
```
save_workflow_version
```
(persist a proposal; always human-approved), then
```
quote_workflow_run
```
and
```
launch_workflow_run
```
.
CLI agent
postplus workflow
commands - the same verbs, results as JSON, requiring a logged-in session (
```
postplus auth login
```
):
```
postplus workflow list
```
/
```
show
```
and
```
runs
```
/
```
run-show
```
(read),
```
create
```
,
```
propose
```
(=
```
propose_edit
```
: preview + validate, persists nothing),
```
save
```
(=
```
save_workflow_version
```
: persist a new immutable version),
```
quote
```
, and
```
launch
```
. Structural edits go in a
```
--operations
```
JSON array of the same
```
add_node
```
/
```
update_node
```
/
```
remove_node
```
/
```
connect_nodes
```
ops. Run
```
postplus workflow help
```
for exact flags.

The two surfaces map one-to-one onto the same hosted verbs and the same validation;

propose

mirrors

propose_edit

save

mirrors

save_workflow_version

quote

mirrors

quote_workflow_run

launch

mirrors

launch_workflow_run

Boundaries:

Resolve ids, never guess them; ambiguous name matches go back to the user as candidates.
Never hand-write a full definition document. Build structure through create plus edit operations and let the server validate; if
```
validation.ok=false
```
, fix the listed errors and re-propose. The server never silently repairs.
Persist only on human approval (
```
save_workflow_version
```
/
```
postplus workflow save
```
), and never launch without explicit user approval - quote first, report the reserved credits, and on the CLI pass the quoted
```
reservedMillicredits
```
as
```
--max-reserved-millicredits
```
together with
```
--confirm
```
(launch refuses to run without both).
If NEITHER surface is available in the current session (no workspace assistant tools and no logged-in CLI), do not fabricate a definition or any substitute file format. Say workflow authoring needs the PostPlus workspace assistant or a logged-in
```
postplus
```
CLI, and route one-off render requests to
```
video-batch-runner
```
.

工作流创作是一项经过服务器验证的功能，有两个入口。使用当前会话提供的任一入口——切勿混合使用，也切勿绕过服务器验证：

工作区助手输入工具：
```
workflow_read
```
（通过
```
action=list
```
按名称查找，通过
```
get
```
查看节点/插槽/版本，通过
```
runs_get
```
查看任务运行记录）、
```
workflow_author
```
（创建空白或基于模板的工作流；
```
propose_edit
```
进行结构性编辑——
```
add_node
```
/
```
update_node
```
/
```
remove_node
```
/
```
connect_nodes
```
——返回真实的验证结果但不保存）、
```
apply_workflow_edit_to_canvas
```
（当用户的实时画布已连接时执行相同操作）、
```
save_workflow_version
```
（保存提案；需人工批准），以及
```
quote_workflow_run
```
和
```
launch_workflow_run
```
。
CLI代理
postplus workflow
命令——动词和结果与上述一致，以JSON格式返回，需要已登录会话（
```
postplus auth login
```
）：
```
postplus workflow list
```
/
```
show
```
和
```
runs
```
/
```
run-show
```
（读取）、
```
create
```
、
```
propose
```
（等同于
```
propose_edit
```
：预览+验证，不保存任何内容）、
```
save
```
（等同于
```
save_workflow_version
```
：保存新的不可变版本）、
```
quote
```
和
```
launch
```
。结构性编辑需放入
```
--operations
```
参数的JSON数组中，包含相同的
```
add_node
```
/
```
update_node
```
/
```
remove_node
```
/
```
connect_nodes
```
操作。运行
```
postplus workflow help
```
查看确切参数。

这两个界面完全对应相同的托管动词和验证机制；

propose

对应

propose_edit

，

save

对应

save_workflow_version

，

quote

对应

quote_workflow_run

，

launch

对应

launch_workflow_run

。

边界规则：

确认ID，切勿猜测；名称匹配模糊时，需向用户返回候选选项。
切勿手动编写完整的定义文档。通过创建加编辑操作构建结构，并由服务器验证；如果
```
validation.ok=false
```
，修复列出的错误后重新提交提案。服务器不会静默修复问题。
仅在人工批准后保存（
```
save_workflow_version
```
/
```
postplus workflow save
```
），且未经用户明确批准切勿启动——先报价，报告预留的积分，在CLI中需将报价的
```
reservedMillicredits
```
作为
```
--max-reserved-millicredits
```
参数，并搭配
```
--confirm
```
参数（缺少任一参数，启动操作都会拒绝执行）。
如果当前会话中两个界面均不可用（无工作区助手工具且未登录CLI），切勿编造定义或任何替代文件格式。告知用户工作流创作需要PostPlus工作区助手或已登录的
```
postplus
```
CLI，并将一次性渲染请求转至
```
video-batch-runner
```
。

Workflow Shape Defaults

工作流形态默认设置

A workflow is a node canvas: generation video nodes (one clip each, each holding its own model-facing prompt plus duration / aspect ratio / resolution / audio knobs in node config) and an optional single assembly node with ordered clip input slots.
With an assembly node, every generation video node must connect into an assembly slot; slot order is cut order. Without one, each node delivers an independent clip.
Passive reference material (creator image, product image, voice sample) is bound as labeled workflow assets referenced from prompts by
```
<<<label>>>
```
; script text can enter as an asset node. Assets do not execute.
Knob values come from the node's selected generation capability; canvas validation and the launch gate are authoritative. Typical UGC defaults:
```
9:16
```
,
```
720p
```
,
```
4-15
```
seconds per clip, audio generation on.
Multi-version output = launching several instances, not duplicating nodes. Name workflows
```
{case-or-brand}-{script-feature}-{duration}s-v{version}
```
unless the user names one.

工作流是一个节点画布：包含生成视频节点（每个节点对应一个片段，每个节点在配置中包含面向模型的提示以及时长/宽高比/分辨率/音频旋钮），以及一个可选的单装配节点，带有有序的片段输入插槽。
如果有装配节点，每个生成视频节点必须连接到一个装配插槽；插槽顺序即为剪辑顺序。如果没有装配节点，每个节点输出独立的片段。
被动参考素材（创作者图片、产品图片、语音样本）作为带标签的工作流资产绑定，在提示中通过
```
<<<label>>>
```
引用；脚本文本可作为资产节点导入。资产不会执行。
旋钮值来自节点所选的生成功能；画布验证和启动网关具有权威性。典型UGC默认值：
```
9:16
```
、
```
720p
```
、每个片段4-15秒、开启音频生成。
多版本输出=启动多个实例，而非复制节点。工作流命名格式为
```
{case-or-brand}-{script-feature}-{duration}s-v{version}
```
，除非用户指定名称。

Decision Router

决策路由

Apply the first matching route:

If	Then
User asks to review, audit, QA, validate, approve, or inspect a workflow	Run L2
Sensitive claims, exact preservation, conflicting facts, rendered-output failures, or batch work	Run L2
User asks to review logic/cut design first, or a complex narrative ad has unlocked hook, pain beat, product entrance, proof/value beat, CTA, or post-production boundary	Return a short cut plan for approval before building
User provides or asks for scripts, 人物图, product 图, reference assets, or per-clip references	Use Script-Locked Asset Mode
Visible product has no product image	Ask by default; generate product imagery only when explicitly permitted
Missing facts would change claims, references, workflow behavior, or naming	Ask the smallest necessary question
Otherwise	Normal mode: read `references/prompt-quality.md` , write clip prompts, build with `workflow_author` , run L0/L1

应用第一个匹配的路由：

条件	操作
用户要求查看、审核、QA、验证、批准或检查工作流	执行L2审核
涉及敏感声明、精确保留要求、事实冲突、渲染输出失败或批量任务	执行L2审核
用户要求先查看逻辑/剪辑设计，或复杂叙事广告包含明确的钩子、痛点环节、产品出场、证明/价值环节、CTA或后期制作边界	返回简短的剪辑计划，待批准后再构建工作流
用户提供或索要脚本、人物图、产品图、参考资产或单片段参考资料	使用脚本锁定资产模式
可见产品无产品图片	默认询问用户；仅在明确允许时生成产品图像
缺少信息会改变声明、参考资料、工作流行为或命名	提出必要的最小问题
其他情况	正常模式：阅读 `references/prompt-quality.md` ，编写片段提示，使用 `workflow_author` 构建，执行L0/L1检查

Three-Judge Frame

三重评判框架

Judge every workflow with three lenses; the rules below only serve these judgments.

Advertiser Judge: product appearance, claims, usage, references, and compliance boundaries match the task-local facts an advertiser would trust.
Viewer Judge: the clip starts from a believable human moment, holds attention, and lands a natural CTA - not a feature announcement.
Generator Judge: downstream models can execute the prompt literally from concrete visible details, clear reference duties, isolated clip scope, and positive scene construction.

从三个维度评判每个工作流；以下规则仅服务于这些评判：

广告主维度：产品外观、声明、使用场景、参考资料和合规边界符合广告主信任的任务本地事实。
观众维度：片段从可信的人类场景切入，保持注意力，自然传递CTA——而非功能公告。
生成器维度：下游模型可根据具体可见细节、明确的参考任务、独立的片段范围和积极的场景构建，按字面执行提示。

Sidecar Guardrail Principle

辅助护栏原则

Guardrail diagnoses, prompt compiles. Business-experience guardrails are sidecar audit checks, not model-facing prompt language. Use them to spot risks (static pacing, audio voids, spatial drift, entity leakage), then translate the smallest truth-preserving repair into positive scene carriers: action, sound source, movement path, prop handling, camera behavior, or a spoken line. Never paste guardrail categories, audit rationale, or banned-term walls into generation prompts.

护栏诊断，提示编译。业务体验护栏是辅助审核检查，而非面向模型的提示语言。用它们识别风险（静态节奏、音频空白、空间偏移、实体泄露），然后将最小的保真修复转化为积极的场景载体：动作、声源、移动路径、道具操作、摄像机行为或台词。切勿将护栏类别、审核理由或禁用术语墙粘贴到生成提示中。

Workflow-First Contract

工作流优先契约

Lock only the needed context: user intent, product facts, claim limits, references, duration, exact hook/VO to preserve, and stated exclusions. If narrative logic is the gap, lock a compact cut plan first (hook, pain beat, product entrance, proof/value beat, CTA, post-production boundary).
Prompt after reading
```
references/prompt-quality.md
```
. Resolve
```
do not/no/avoid/不要
```
into positive scene carriers; bind every reference to a job, scope, and boundary. Duration, aspect ratio, resolution, and audio generation belong in node config, never restated as prompt text.
Build through
```
workflow_author
```
(or
```
postplus workflow create
```
/
```
propose
```
): create (blank or template), then propose the node/edge/config structure with edit operations.
Validate with the returned verdict plus the internal quality gate. Fix truth-preserving issues directly; block only for missing facts that would change claims or behavior.

Persist and launch on human approval only:

save_workflow_version

(or

postplus workflow save

), then

quote_workflow_run

(or

postplus workflow quote

), report the reserved cost in credits (millicredits ÷ 1000), and

launch_workflow_run

(or

postplus workflow launch ... --confirm

) after explicit approval - pass the quote's

reservedMillicredits

maxTotalReservedMillicredits

--max-reserved-millicredits

and the exact workflow name as

workflowTitle

--title

Handoff workflow name and id, saved version, validation status, quote/launch status, and meaningful blockers or residual risks. Runs are asynchronous with human review gates - report state and the workflow page link; do not poll a run to completion.

锁定必要的上下文：用户意图、产品事实、声明限制、参考资料、时长、需保留的精确钩子/旁白，以及明确的排除项。如果叙事逻辑存在缺口，先锁定紧凑的剪辑计划（钩子、痛点环节、产品出场、证明/价值环节、CTA、后期制作边界）。
编写提示前阅读
```
references/prompt-quality.md
```
。将
```
do not/no/avoid/不要
```
转化为积极的场景载体；为每个参考资料绑定任务、范围和边界。时长、宽高比、分辨率和音频生成属于节点配置，切勿在提示文本中重复。
构建通过
```
workflow_author
```
（或
```
postplus workflow create
```
/
```
propose
```
）：创建（空白或模板），然后通过编辑操作提交节点/连接/配置结构提案。
验证结合返回的结果和内部质量检查。直接修复保真问题；仅当缺少会改变声明或行为的信息时才阻止操作。
仅在人工批准后保存并启动：执行
```
save_workflow_version
```
（或
```
postplus workflow save
```
），然后执行
```
quote_workflow_run
```
（或
```
postplus workflow quote
```
），报告预留的积分（毫积分÷1000），获得明确批准后执行
```
launch_workflow_run
```
（或
```
postplus workflow launch ... --confirm
```
）——将报价的
```
reservedMillicredits
```
作为
```
maxTotalReservedMillicredits
```
/
```
--max-reserved-millicredits
```
参数，将确切的工作流名称作为
```
workflowTitle
```
/
```
--title
```
参数。
交接工作流名称和ID、保存的版本号、验证状态、报价/启动状态以及有意义的阻塞因素或剩余风险。任务运行是异步的，带有人工审核关卡——报告状态和工作流页面链接；无需轮询任务直至完成。

Script-Locked Asset Mode

脚本锁定资产模式

Trigger when the user plans a script first, mentions 人物图 / product 图 / reference images, provides product images without a creator image, or wants per-clip references. Lock the script (hook, lines, CTA, total duration, product actions, claim limits) before graph construction; lock

creator_identity

product_reference

voice_reference

, and

style_reference

as labeled assets; plan clips with duration, beat, product presence, and scene-reference needs. Prefer the graph

creator identity image -> per-clip scene image -> per-clip generation video -> assembly

. Read

references/asset-locked-workflow.md

only when this mode triggers, and follow its build order.

当用户优先规划脚本、提及人物图/产品图/参考图片、提供产品图片但无创作者图片，或需要单片段参考资料时触发此模式。在构建图表前锁定脚本（钩子、台词、CTA、总时长、产品动作、声明限制）；将

creator_identity

、

product_reference

、

voice_reference

和

style_reference

作为带标签的资产锁定；规划片段的时长、环节、产品出现情况和场景参考需求。优先采用图表结构

创作者身份图片 -> 单片段场景图片 -> 单片段生成视频 -> 装配

。仅在此模式触发时阅读

references/asset-locked-workflow.md

，并遵循其构建顺序。

Prompt Contract

提示契约

Write only model-facing generation instructions; keep internal constraints, scene carriers, and the final model prompt as separate layers. Every generation node prompt needs: reference bindings with job/scope/boundary, real handheld UGC style and setting, one speech posture (

on-camera spoken

voiceover

silent

), timestamped cuts with concrete actions and cut-bound spoken text, audio rules, and visual constraints with positive replacements for unwanted UI/text/logos. Keep claims inside task-local facts; keep clip references isolated; use spoken density (typical

2.4-2.8

words/second; yapping

3.0-3.8

only on request; trust-sensitive

2.2-2.6

) as the duration rule, and require a new beat every

1-2s

for high-density briefs. Every cut must be physically executable by one plausible camera operator. Full doctrine, examples, and camera grammar live in

references/prompt-quality.md

If the user wants external music, later sound design, or a silent workflow, turn the node's audio generation off and remove spoken dialogue and voice references from prompts.

仅编写面向模型的生成指令；将内部约束、场景载体和最终模型提示分为独立层级。每个生成节点的提示需要：带有任务/范围/边界的参考绑定、真实手持UGC风格和场景、一种语音姿态（

on-camera spoken

voiceover

silent

）、带时间戳的剪辑（包含具体动作和剪辑边界的台词）、音频规则，以及带有不想要的UI/文本/标志的积极替代方案的视觉约束。确保声明符合任务本地事实；确保片段参考独立；使用说话密度（典型值2.4-2.8词/秒；仅在请求时使用3.0-3.8词/秒的快节奏；敏感场景使用2.2-2.6词/秒）作为时长规则，高密度需求下每1-2秒需要一个新环节。每个剪辑必须可由一个合理的摄像师实际执行。完整规则、示例和摄像语法见

references/prompt-quality.md

。

如果用户需要外部音乐、后期音效设计或静音工作流，关闭节点的音频生成，并从提示中移除口头对话和语音参考。

Internal Quality Gate

内部质量检查

L0 hard check (always, from the propose result and

workflow_read

get): validation verdict clean; with an assembly node every generation node feeds an ordered slot; every generation node has intended prompt + duration / aspect / resolution / audio config and reference bindings; name-embedded duration matches total clip duration; prompts carry no post-production leakage, negative-composition phrasing, or unsupported medical/legal/guarantee claims. Treat scan hits as signals; fix only what affects launchability, factuality, or generation quality.

L1 self review (default before handoff): Truth - latest intent and constraints preserved, no invented entities, features, claims, or UI. Watchability - first

0-5s

creates a viewer question; product entry is believable; payoff and CTA read naturally. Executability - prompts stay model-facing; references isolated with explicit duties; negatives compiled into positive carriers; cuts physically plausible. Apply small truth-preserving fixes and re-run L0.

L2 deep review (auto-trigger; never ask): sensitive claims; exact-preservation requirements; edits made outside the propose/validate loop; L0 contradiction signals; conflicting facts or references; batch work; 30s+ multi-clip continuity of person/product/setting/story; multi-area scene topology; reported rendered-output issues; or an explicit review request. Read the full definition via

workflow_read

get, then

references/ugc-workflow-audit-rubric.md

; add

references/ugc-sidecar-guardrails.md

for rhythm/audio/spatial/leakage triggers, treating sidecar findings as diagnostics unless they expose factual, product, compliance, or reference-binding issues. For assembled videos, review cross-clip continuity of identity, props, locations, dialogue flow, and post-production leakage. Use a fresh review context per workflow when independent review helps; reviewers only review, the parent owns edits via propose/apply operations. Stop after pass, blocker, or 3 review-revision cycles; block and ask the smallest question only when missing facts would change claims.

L0硬性检查（始终执行，来自提案结果和

workflow_read

的

get

操作）：验证结果无问题；如果有装配节点，每个生成节点都连接到有序插槽；每个生成节点有预期的提示+时长/宽高比/分辨率/音频配置和参考绑定；名称中嵌入的时长与总片段时长匹配；提示无后期制作泄露、负面构图表述或未经许可的医疗/法律/担保声明。将扫描结果视为信号；仅修复影响启动可行性、真实性或生成质量的问题。

L1自我检查（交接前默认执行）：真实性——保留最新意图和约束，无虚构实体、功能、声明或UI。观赏性——前0-5秒引发观众兴趣；产品出场自然；收益点和CTA传递自然。可执行性——提示仅面向模型；参考资料带有明确任务且独立；负面要求转化为积极载体；剪辑在物理上可行。进行小的保真修复后重新执行L0检查。

L2深度检查（自动触发；无需询问）：涉及敏感声明；有精确保留要求；在提案/验证循环外进行的编辑；L0矛盾信号；事实或参考资料冲突；批量任务；30秒以上的多片段人物/产品/场景/故事连续性；多区域场景拓扑；报告的渲染输出问题；或明确的审核请求。通过

workflow_read

的

get

操作读取完整定义，然后阅读

references/ugc-workflow-audit-rubric.md

；如果涉及节奏/音频/空间/泄露问题，添加

references/ugc-sidecar-guardrails.md

，将辅助检查结果视为诊断信息，除非它们暴露真实性、产品、合规或参考绑定问题。对于装配好的视频，检查跨片段的身份、道具、场景、对话流和后期制作泄露的连续性。当独立审核有帮助时，为每个工作流使用全新的审核上下文；审核人员仅负责审核，由创建者通过提案/应用操作进行编辑。通过、阻塞或经过3次审核-修订循环后停止；仅当缺少会改变声明的信息时才阻塞并提出必要的最小问题。

Cost And Launch Discipline

成本与启动规范

Launching spends real credits: always quote first, report the reserved credits, and never launch without explicit user approval. On the CLI,

postplus workflow launch

refuses to run without

--confirm

and an acknowledged

--max-reserved-millicredits

ceiling; the server re-quotes and aborts if the fresh reservation exceeds it, so the confirmed cost bound stays binding. Runs pause at human review gates (per-clip approval, final review) - report run state and the workflow page link instead of sitting in a polling loop.

启动任务会消耗真实积分：始终先报价，报告预留积分，未经用户明确批准切勿启动。在CLI中，

postplus workflow launch

需要

--confirm

参数和已确认的

--max-reserved-millicredits

上限才会执行；如果新的预留积分超过该上限，服务器会重新报价并中止操作，因此确认的成本上限始终有效。任务会在人工审核关卡（单片段批准、最终审核）暂停——报告任务状态和工作流页面链接，而非持续轮询。

Handoff

交接

Report the workflow name and id, saved version number, validation status, quote or launch status with reserved credits, and blocked facts or meaningful residual risks. Do not expose internal QA steps, review cycles, or audit rationale unless the user asks.

报告工作流名称和ID、保存的版本号、验证状态、报价或启动状态及预留积分，以及阻塞因素或有意义的剩余风险。除非用户询问，否则不要暴露内部QA步骤、审核周期或审核理由。