这页照实写我们怎么选品：想法怎么变成候选需求，同一条流水线怎么把它们排成机会，三个深度各能得出什么结论。每个说法都来自我们的方法论正本；能公开的数据和实验记录，你可以自己核对、自己下判断。This page sets out how we pick what to build: how an idea becomes candidate needs, how one pipeline ranks them into opportunities, and what each of the three depths can conclude. Every statement comes from our methodology; where the data and experiment records are public, you can check them and judge for yourself.
用户带着一个想法来（常常拿已有产品举例），我们把它变成一组候选需求，用同一条流水线判断哪些是值得做的机会，每个结论都附可复查的数据和实验记录，能回测的用历史数据检验它过去准不准。A user arrives with an idea, often pointing at an existing product. We turn it into a set of candidate needs, use one pipeline to judge which of them are opportunities worth building, attach checkable data and an experiment record to every conclusion, and where a backtest is possible, test with historical data whether it was right in the past.
| 对象Object | 是什么What it is | 例子Example |
|---|---|---|
| 需求Need | 用户要完成的一件事；能互相替代的产品属于同一个需求。One thing a user is trying to get done; products that can substitute for each other belong to the same need. | “在网页侧边栏和 AI 聊天”“Chat with AI in a browser side panel” |
| 产品Product | 某个平台上已有的一个具体解决方案：Chrome / Edge / Firefox 扩展、Android / iOS App，以后还有网站工具。页面按平台显示“扩展”“App”，统称“产品”。An existing concrete solution on a platform: a Chrome / Edge / Firefox extension, an Android / iOS app, and later web tools. The page labels them “extension” or “app” by platform; collectively we call them products. | Claude、ChatHub、Sidebar HeroClaude, ChatHub, Sidebar Hero |
| 机会Opportunity | 值得做的需求 + 切口（凭什么能做得更好）+ 建议形态。A need worth building + an angle (why we can do it better) + a suggested product shape. | “AI 侧边栏：做本地无账号、按次付费的版本”“AI side panel: a local, no-account, pay-per-use version” |
结果只有一种：机会（需求）。产品是入口和证据，不是另一种结果。每个产品归到它的主需求。There is only one kind of result: an opportunity (a need). Products are the entry point and the evidence, not another kind of result. Every product is grouped under its main need.
入口：① 信号：为什么这里有缺口（7 个） ② 产品条件：如“付费且近 30 天在涨” ③ 以产品为例：“像 Claude 这种”Entry points: (1) signals: why there is a gap here (7 of them) · (2) product conditions, e.g. “paid and growing over the last 30 days” · (3) start from a product, e.g. “something like Claude”
流水线：候选需求 → 筛选 → 评分项排序 → 形态 → 研判 → 组批Pipeline: candidate needs → screening → ranking by scoring items → product shape → verdict → batching
| 深度Depth | 做什么What it does | 花费Cost | 结论怎么标How the conclusion is labelled |
|---|---|---|---|
| 查Look | 只读数据：筛选、排序、需求卡片、趋势。Read-only data: screening, ranking, need cards, trends. | 免费、即时Free, immediate | “数据事实”（附数据日期和精度说明）“Data fact”, with the data date and a note on precision |
| 评Assess | 加评分项、形态、联网研判。Adds scoring items, product shape and an online verdict. | 研判计额度Verdicts draw on the quota | “基于当前数据的判断，未经历史回测验证”“A judgement based on current data, not yet validated by a historical backtest” |
| 验Validate | 用历史数据回放这套评估标准，看过去准不准。Replays the evaluation criteria on historical data to see whether they were right in the past. | 补数据计额度Filling in data draws on the quota | “有效 / 观察 / 无效 / 样本不足”，附提升倍数和区间“Effective / observed / not effective / insufficient sample”, with the lift and interval |
| 选品：“这个方向值不值得做”What to build: “is this direction worth doing?” | 运营：“这个产品还值不值得继续投入”What to keep: “is this product still worth investing in?” | |
|---|---|---|
| 对象Object | 需求，以及后来者有没有机会A need, and whether a newcomer has a chance | 一个具体产品，以及它自己的走势One specific product and its own trajectory |
| 数据和工具Data and tools | 同一批The same set | 同一批The same set |
| 回测时什么算“对”What counts as “right” in a backtest | 12 个月内这个需求冒出 ≥1 万用户的非克隆后来者Within 12 months the need produces a non-clone newcomer with at least 10,000 users | 之后 6–12 个月这个产品的用户和评分涨跌、有没有被竞品超过（口径待定，定之前只能“评”不能“验”）How the product’s users and ratings move over the next 6–12 months, and whether a competitor overtakes it (definition pending; until it is fixed, this can only be assessed, not validated) |
两类问题都可以回答；运营类结论在口径定好、回测过之前，一律标“未经验证的判断”。Both kinds of question can be answered; before the definition is fixed and backtested, every answer to the second kind is labelled “an unvalidated judgement”.
| 层Layer | 例子Examples | 谁能定义Who can define it |
|---|---|---|
| 原子Atomic | 用户数、评分、评分数User count, rating, number of ratings | 平台The platform |
| 属性Attribute | 商店、类目、收费方式、开发者、上架日期、所属需求Store, category, pricing model, developer, release date, main need | 平台The platform |
| 派生Derived | 近 30 天用户净增长、新增评分数、上架天数Net user growth over the last 30 days, new ratings, days since release | 用户可在登记表规则内组合Users can combine them within the registry rules |
| 复合Composite | 增长率、需求合计用户、信号、评分项Growth rate, total users in a need, signals, scoring items | 用户可写白名单公式Users can write whitelisted formulas |
| 判断Judgement | 收费模式、差评主题、产品优势Pricing model, themes in negative reviews, product advantage | 平台；用户自定义按条计费（以后）The platform; user-defined judgements are charged per item (later) |
每个指标都登记精度（例如用户数 ≥1,000 时是取整显示值）、最早日期、能否回放、成本和版本。“属性”在早先的设计里叫“维度”，为免和原“6 维”混淆改了名。Every metric records its precision (for example, user counts of 1,000 or more are rounded display values), earliest date, whether it can be replayed, its cost and its version. What is now called “attribute” was called “dimension” in earlier designs; it was renamed to avoid confusion with the former “6 dimensions”.
评估标准把上面的选择写成一份可保存、可复制、可对比、可回测的文件：用哪种入口、范围和需求条件怎么设、六项评分项和可信度、建议形态、研判结论与前提。同一份评估标准换一个时间点回放，才能比较“过去准不准”。The evaluation criteria write those choices down as a file that can be saved, copied, compared and replayed: which entry point, how range and need conditions are set, the six scoring items and their confidence, the suggested shape, and the verdict with its preconditions. Replaying the same criteria at a different point in time is what makes “was it right in the past?” comparable.
我们自己的偏好是一份“默认评估标准”，公开给别人参考。你可以用自己的条件另写一份，再拿同一套回放方法比较。Our own preference is the “default evaluation criteria”, published here for reference. You can write your own with your own conditions and compare it using the same replay method.
这套评估标准过去准不准，由回测来回答：读者版结果在 /backtest/；研判回测的公开文章在 /articles/e29/（已公开可能性分数 AUC 0.88）。Whether these evaluation criteria were right in the past is what the backtest answers: the reader edition is at /backtest/, and the public write-up of the verdict backtest is at /articles/e29/ (it publishes an AUC of 0.88 for the likelihood score).
评估标准回放约 1.8 倍——80 个历史需求里命中 28 个，随机同规模期望 15.2 个，90% 区间 1.2–2.4。Across 80 historical needs, 28 surfaced by our evaluation criteria later succeeded — about 1.8× the 15.2 expected by chance (90% interval 1.2–2.4).
边界也照实说：这一轮回测只回放了信号和范围，评分项、形态和研判没有进入回放名单；研判这一步在回测里没有显示增量。所以对外只报评估标准名单的倍数，不把研判说成准确率或预测率，也不把 1.8 倍说成研判的效果。The limits are stated too: this replay covered signals and range only; scoring items, product shape and the verdict were not part of the replayed list, and the verdict step showed no added value in the backtest. So we report only the multiple for the evaluation-criteria list: we do not present the verdict as an accuracy or prediction rate, and we do not present 1.8× as the verdict’s effect.