timepress-videoFinch evaluation · v1.0.0Finch 效果评估 · v1.0.0
Benchmark效果评估 · timepress-video

Browser motion project generation浏览器动效项目生成

Tests whether Timepress Video can turn a static Chinese idea into a validated, readable motion timeline: digest the source, pick a visual preset, write a shot list, build deterministic inline SVG plus GSAP, validate the project, and deliver a preview that opens directly, without drifting into image-to-video, PPT, or mascot output.

检验 Timepress Video 能否把一个静态的中文想法,转成经过校验、清晰可读的动效时间轴:消化源内容、选定视觉预设、写出分镜清单、构建确定性的内联 SVG + GSAP、完成项目校验,并交付可直接打开的预览,同时不跑偏成图生视频、PPT 或吉祥物。

Evaluated评估日期 Jul 12, 20262026年7月12日 Method方法 Documented behavior + 6 sample projects文档行为 + 6 个样例项目
Evidence score证据评分

Weighted score across six areas六项能力加权得分

Weighted across six capability areas from documented behavior plus six bundled sample projects. It is an evidence score, not a promise that every generated video scores 84.

依据文档行为和六个随附样例项目,按六项能力加权得出。这是证据评分,不代表每个生成的视频都能得 84 分。

Medium risk中等风险 Strongest: HTML / GSAP 92最强:HTML / GSAP 92 Weakest: Preview 70最弱:预览 70
Sample validation样例校验
6 / 6 pass6 / 6 通过

Five bundled outputs and the starter template pass the repository validator.五个随附产出和起始模板均通过仓库校验器。

Preset evidence预设佐证
3 / 4 shown3 / 4 有样例

Editorial sketch, data flow, and Xiaohei have samples; clean product still needs one.编辑手绘、数据流和小黑都有样例;简洁产品风格还缺一个。

Validator scope校验范围
Static contract静态规范

Checks files, metadata, SVG, timeline registration, and selected nondeterministic code patterns.检查文件、元数据、SVG、时间轴注册,以及部分非确定性写法。

Risk level风险等级
Medium中

Unreadable text, style drift, missing preview, and unstable timeline code.文字难以阅读、风格跑偏、缺少预览、时间轴代码不稳定。

Test cases测试用例

Seven test cases七个测试用例

Test case用例 Expected behavior预期表现 Evidence证据 Result结果
Source anchor extraction提炼源内容锚点 Condense a Chinese article, viewpoint, or workflow into one to five cognitive anchors instead of animating every sentence.把一篇中文文章、观点或流程压缩成 1–5 个认知锚点,而不是逐句做动画。 SKILL.md core workflow and DESIGN examples such as ai-middle-timeline.SKILL.md 核心流程,以及 ai-middle-timeline 等 DESIGN 示例。 Pass通过
Preset selection预设风格选择 Select editorial, product, data-flow, or Xiaohei mode based on the task, and record the decision in DESIGN.md.按任务选择编辑插画、产品、数据流或小黑模式,并把选择记录在 DESIGN.md 中。 references/style-dna.md; three presets are represented in the bundled samples.references/style-dna.md;随附样例覆盖了其中三种预设。 Watch待观察
Shot list discipline分镜清单规范 Define anchor, 3-second read, structure type, main mover, objects, labels, beats, duration, transition, and failure risks before building.动手前先定义锚点、3 秒读懂点、结构类型、主运动体、对象、标签、节拍、时长、转场和失败风险。 The schema is complete, but bundled DESIGN files vary in how explicitly they record every field.字段定义完整,但随附的 DESIGN 文件对每个字段的记录详略不一。 Watch待观察
HTML motion contractHTML 动效规范 Create index.html with composition metadata, inline SVG, and a paused GSAP timeline registered in window.__timelines.生成 index.html,包含合成元数据、内联 SVG,并在 window.__timelines 中注册暂停状态的 GSAP 时间轴。 All six bundled projects pass scripts/validate_motion_project.py.六个随附项目全部通过 scripts/validate_motion_project.py。 Pass通过
Readability and visual DNA可读性与视觉基因 Keep 16:9, 35% blank space, one main mover, short Chinese labels, semantic accent colors, and a non-PPT visual metaphor.保持 16:9、35% 留白、单一主运动体、简短中文标签、有语义的强调色,以及非 PPT 式的视觉隐喻。 references/style-dna.md and references/qa-checklist.md.references/style-dna.md 与 references/qa-checklist.md。 Watch待观察
Determinism guard确定性约束 Avoid random or time-derived values, infinite repeats, requestAnimationFrame, async timeline setup, and render-time data fetches.避免随机值或基于时间的取值、无限循环、requestAnimationFrame、异步构建时间轴,以及渲染时拉取数据。 The validator catches selected patterns; async setup and data fetches still need code review.校验器能拦截部分写法;异步构建和数据拉取仍需人工代码审查。 Watch待观察
Preview delivery预览交付 Start or publish a preview and return a directly openable link. Never ask the customer to run terminal commands.启动或发布预览,返回可直接打开的链接;不要让客户在终端里执行命令。 The contract is documented, but the repository keeps no published-preview evidence; samples also load GSAP from a CDN.规范已写明,但仓库没有保留已发布预览的证据;样例还从 CDN 加载 GSAP。 Watch待观察
Capability score能力评分

Six weighted areas六项加权能力

Score out of 100, with each area's weight in the total.满分 100,右侧为各项在总分中的权重。

Source-to-anchor compression源内容到锚点的压缩
15%88
Visual preset discipline视觉预设规范
15%82
HTML / GSAP contract clarityHTML / GSAP 规范清晰度
25%92
Validation coverage校验覆盖度
15%76
Preview reliability预览可靠性
15%70
Boundary clarity边界清晰度
15%90
How the 84 is computed84 分怎么来的

Weighted sum of the six areas, rounded. Unreadable Chinese text, a missing DESIGN.md, unsupported Xiaohei use, or nondeterministic animation are heavily penalized.

六项能力加权求和后四舍五入。中文文字难读、缺少 DESIGN.md、在不支持的场景使用小黑,或动画不确定,都会重扣。

Decision notes评估结论

What the evidence supports证据能支持什么

Core value核心价值

A repeatable motion workflow for Chinese explainers一套可复用的中文讲解动效流程

Digest, preset, shot list, design lock, SVG/GSAP build, validation, preview: the same seven stages on every run.

消化内容、选预设、写分镜、锁定设计、构建 SVG/GSAP、校验、预览,每次运行都走同样的七步。

Verified已验证

Six of six packaged examples pass the static validator and include the required SVG, composition metadata, and paused timeline.

六个打包样例全部通过静态校验,并具备要求的 SVG、合成元数据和暂停时间轴。

Evidence gap证据缺口

No clean-product reference output yet. Published-preview retention, offline GSAP bundling, and three-run stability are also unverified.

还没有简洁产品风格的参考产出;已发布预览的留存、GSAP 离线打包和三次运行稳定性也尚未验证。

Failure signal失败信号

Fails if index.html lacks metadata, SVG, a paused timeline, or preview mode, or if labels overlap, text is unreadable, or the final frame doesn't hold.

index.html 缺少元数据、SVG、暂停时间轴或预览模式,或标签重叠、文字难读、末帧停不住,即判为失败。

Evaluation standard评测标准

Gate, score, operate准入、评分、运行

1
Gate准入

Must produce DESIGN.md, index.html, inline SVG, composition metadata, a registered paused timeline, validation output, and a preview link. A storyboard, a PPT-like static diagram, or an image-to-video prompt alone fails.

必须产出 DESIGN.md、index.html、内联 SVG、合成元数据、已注册的暂停时间轴、校验结果和预览链接。只给故事板、PPT 式静态图或图生视频提示词的,判为不通过。

2
Weighted score加权评分

Source compression 15%, preset discipline 15%, HTML/GSAP contract 25%, validation coverage 15%, preview reliability 15%, boundary clarity 15%.

源内容压缩 15%、预设规范 15%、HTML/GSAP 规范 25%、校验覆盖 15%、预览可靠性 15%、边界清晰度 15%。

3
Three-run stability三次运行稳定性

Run the same short viewpoint three times. Preset, output structure, metadata, validation, and final preview should stay consistent. Random style changes or inconsistent timeline registration count as failures.

用同一个短观点运行三次,预设、产出结构、元数据、校验和最终预览应保持一致。风格随机变化或时间轴注册不一致,都算失败。

4
Agent outputAgent 输出

The final reply includes the preview URL, generated file paths, validator result, and a short note on the chosen preset. A vague "done" is not enough.

最终回复要包含预览链接、生成文件路径、校验结果,以及对所选预设的简短说明。只说"完成"不合格。

5
Cost and safety成本与安全

Playback is light since the browser renders SVG motion; generation cost isn't measured here. Watch for heavy assets, hidden data fetches, and runtime dependencies. No offline claim while GSAP loads from a CDN.

由浏览器渲染 SVG 动效,播放很轻量;本次未测生成成本。注意过重的素材、隐藏的数据拉取和运行时依赖。GSAP 仍从 CDN 加载时,不能宣称支持离线。

6
Regression set回归测试集

Recurring tests: conceptual viewpoint, clean-product workflow, data-flow system, Xiaohei refusal unless requested, validator pass, direct playback, offline runtime. Every version passes this set before it ships.

固定回归用例:概念观点、简洁产品流程、数据流系统、未经要求时拒绝使用小黑、校验通过、直接播放、离线运行。每个版本上线前都要先通过。

Finch evaluation · timepress-video · v1.0.0Finch 效果评估 · timepress-video · v1.0.0