Seven directions the redesign points toward.7 个 方向,沿着重设计继续走下去。
Where these come from
The five solution moves answer the take-home brief. The seven directions here answer the brief’s final question: «If you had more time, what would you explore next?» Each is grounded in a real gap we found during diagnosis, but none was small enough to ship in eight hours.
这些方向是哪来的
5 个 Solution 里的 move 回答了 take-home 的核心问题。下面 7 个方向回答了 brief 最后一问:「如果时间更多,你还想探索 什么?」每一个都基于诊断阶段找到的真实缺口,但都没小到 能在 8 小时内做完。
Close the prediction-to-truth loop把「预测」和「真实数据」连成闭环
Every variant carries a predicted metric lift (Move 04). The next quiet integration to build is the one that compares prediction to outcome.
When a marketer pushes a variant to Meta, the platform begins reporting performance ~48 hours later. The system should auto-ingest that performance, bind it back to the original prediction (variant id, recommendation id, predicted lift), and tune the confidence model.
This is what turns the +38% chip from a guess into a calibrated bet. It’s also what makes Atria sticky — every successful prediction trains the next.
Move 04 给每张变体加了「预测涨幅」。下一步该建的、最不显眼但 最关键的能力是 —— 把「预测」和「真实结果」对比起来。
marketer 把变体推回 Meta 之后,Meta 大概 48 小时后开始回报真实 表现。系统应该自动拉这些真实数据,绑定到原来的预测上(变体 id、建议 id、预测涨幅),然后用它来校准置信度模型。
这就是让「+38%」从「猜测」变成「校准过的赌注」的关键。也是 Atria 留住用户的关键 —— 每一次成功的预测都在训练下一次。
Video iteration MVP视频迭代最小可用版本
Today the redesign exposes a «Brief export» action for video ads (Move 03). That’s the safe minimum. The next step is to handle the parts of video iteration that don’t require full post-production:
- Copy / caption overlay variants on the same video.
- Audio (voiceover, music) replacement on the same cut.
- Hook-time recommendation: «move the product reveal to 0:03 instead of 0:08».
These don’t need a full video pipeline. They need transcript + scene boundary detection + overlay rendering. The full brief export stays as a fallback for true re-cuts.
目前重设计为视频广告提供了一个「导出 brief 给后期」的动作 (Move 03)。这是最保守的兜底。下一步是处理「不需要从头剪 视频」的轻量编辑:
- 给同一条视频换文案叠加 / 字幕变体。
- 给同一条剪辑换配音 / 换背景音乐。
- Hook 时间建议:「把产品露出从 0:08 提前到 0:03」。
这些不需要完整的视频管道。需要的是:转录(把视频里的话转成 文字)+ 镜头切分识别 + 文案叠加渲染。完整的「导出 brief」 继续作为「真的需要重剪」的兜底。
Multi-user review for variants变体的多人审核工作流
Agencies don’t ship from a single account. A strategist generates variants, an account director reviews, the brand client signs off. Today, the Variants page assumes a single decision maker.
Each variant card should support: comment threads (Linear-style inline), tags (approved / rejected / needs revision), and a shareable read-only link with the source context (which Radar row, which recommendation).
The selection-as-primary pattern from Move 05 extends naturally: select 4 variants → «Send for review» → name reviewers → notification in their Tasks panel.
代理公司不是一个人决策。Strategist(策略师)出变体,Account Director(客户总监)审核,品牌客户最终签字。但现在的 Variants 页面假设只有一个决策者。
每张变体卡应该支持:评论区(像 Linear 那样内嵌的评论线索)、 标签(已通过 / 已驳回 / 需要修改)、以及一个带上下文的只读 分享链接(显示这是 Radar 哪一行、哪条建议触发的)。
Move 05 的「选中即操作」逻辑可以自然延伸:选中 4 张变体 → 「发送审核」→ 指定审核人 → 审核人的 Tasks 面板收到通知。
Brand profile as a first-class object把 Brand profile 做成一等公民
Brand profile (brand identity record) is currently a prerequisite the user trips over inside the Add Context modal. The redesign smooths that trip, but the deeper move is to treat Brand as a versioned, sharable object:
- Brand voice, palette, typography, tone — all live in a structured form.
- Each iteration logs which brand version it used.
- Strategists can fork the brand for «what-if» tests (e.g., what if we used cooler palettes for this seasonal campaign).
This is also the entry point for a true asset library where past variants become reusable kit.
Brand profile(品牌档案,记录这个品牌的视觉、口吻、定位等) 目前是用户在 Add Context 弹窗里会绊到的前置条件。重设计把这个 绊脚石平了。但更深一层的动作是:把 Brand 做成可版本化、可分享 的一等对象:
- 品牌口吻、配色、字体、调性 —— 全部存进结构化的字段里。
- 每次迭代记录用了哪个品牌版本。
- 策略师可以 fork 一个品牌做「假如」测试(比如:这个节日 活动如果用冷色调会怎么样)。
这也是做真正的素材库的入口 —— 让过去的变体变成可以复用的 资产包。
Feedback that actually changes something真的能产生影响的反馈机制
«Was this helpful?» is the laziest feedback pattern in software. We didn’t fix it in the mockup; we should.
Two upgrades:
- Bind feedback to the (recommendation id, ad id) pair. Five «not helpful» on the same recommendation type for the same ad cluster should auto-downrank that type for similar ads.
- Make feedback a private signal AND a public one. Private: trains the model. Public: shows the user that their feedback moves the needle («3 strategists found this less useful — we de-prioritized similar recs for your account»).
「这条有用吗?」是软件里最偷懒的反馈模式。原型里没修这个, 应该修。
两个改动:
- 把反馈绑定到「建议 id + 广告 id」组合上。同一类建议在同类 广告上被打了 5 次「没用」之后,该类建议在这类广告上自动 降权。
- 反馈既是「私有信号」也是「公开信号」。私有:训练模型。 公开:让用户看见自己的反馈在改变什么(「3 个策略师都觉得 这条建议没用 —— 我们已经在你账户里给同类建议降低了优先级」)。
Role-based presets instead of column customization按角色的预设视图,代替「自定义列」
«Customize columns» is currently a power-user feature for the wrong user. The named user — creative strategist — needs a few well-tuned views, not a column-by-column custom mode.
Three default views: Strategist (creative-focused columns), Analyst (metric-heavy), Brand manager (paid spend + ROAS). Switching is one click. The full Customize stays as an escape hatch in settings.
「自定义列」目前是个给错了用户的高级功能。Atria 的目标用户 —— creative strategist(创意策略师)—— 需要的不是「逐列自定义」 的强大,而是几个调校好的视图。
三个默认视图:Strategist(策略师视图,创意相关列) / Analyst(分析师视图,指标密集) / Brand manager(品牌经理视图, 花费 + ROAS)。切换一键完成。完整的自定义功能保留在设置页 作为兜底。
Platform-aware variants懂投放平台规格的变体
Meta, Google, TikTok, Pinterest all have different image specs (1:1, 4:5, 16:9, story). Today the user generates 4 variants per group and figures out platform fit later. The next move is to make platform-aware variants explicit:
- Each variant carries its valid platform set.
- A «for» control lets users generate platform-specific batches.
- Push-to-platform routes only the compatible variants.
This makes the «push» button do real work, not just open the ad manager in a new tab.
Meta、Google、TikTok、Pinterest 各自有不同的图片规格(1:1、 4:5、16:9、竖屏故事流)。现在用户出 4 张变体之后,要自己判断 哪张能投到哪个平台。下一步是让「平台规格」显性化:
- 每张变体带上「适配的平台」标签。
- 顶部一个「为...生成」的下拉,可以指定平台批量生成。
- 「推回投放」按钮只把适配的变体推到对应平台。
这样「推回投放」这个动作就是真的在做事,而不是只是给你开个 后台标签页让你自己复制粘贴。
Closing
The take-home is bounded. What this redesign argues for, at a higher level, is a repositioning: Atria stops being «an analytics tool with a generator bolted on» and becomes «an AI copilot for creative decisions». The five moves we shipped are the structural shift; the seven directions above are how that shift compounds into a moat.
If a single picture summarizes it: today Radar reads like a leaderboard. We want it to read like a todo list.
收尾
take-home 是有边界的。这个重设计在更高层面想说的是 —— Atria 应该重新定位:从「分析工具 + 一个外挂的生成器」变成「为 创意决策服务的 AI 副驾」。原型里跑通的 5 个 move 是结构性 的转变;上面 7 个方向是「这个转变怎么积累成护城河」。
如果用一张图总结:今天的 Radar 读起来像排行榜。我们要让它读起来像待办清单。