verify-then-advise

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Verify, Then Advise

先验证，再给出建议

Some advice is wrong the moment you give it, and nothing in the answer reveals it. The structure is sound, the reasoning holds, the tone is confident — and the central fact retired six weeks ago.

This skill exists for recommendations that rest on facts you do not own: what a vendor still sells, what an exam still certifies, what a market actually pays for. Those facts have a half-life, and your memory of them does not come with an expiry date attached.

The governing idea, borrowed from the skills-lifecycle literature: a stale skill is worse than no skill, because it gives confident but outdated instructions. Stale model knowledge fails the same way, and is harder to notice, because there is no file to check the date on.

有些建议在你说出的那一刻就是错误的，但答案本身却毫无征兆。结构合理、逻辑通顺、语气自信——但核心事实早在六周前就已失效。

本技能专为依赖非自有事实的建议场景而设计：供应商仍在销售的产品、考试仍认可的资质、市场实际支付的薪资等。这些事实有半衰期，而你对它们的记忆并没有附带过期日期。

其核心理念借鉴自技能生命周期相关文献：过时的技能不如没有技能，因为它会给出自信但过时的指导。过时的模型知识也会以同样方式失效，而且更难被察觉，因为没有文件可以查看日期。

When to invoke

何时调用

"Which certification should I take?" · "Is X still supported?" · "What's the market for Y?" · "Should we adopt Z?" · "What are the trends in…?" · "Where should my career go in five years?" · any roadmap spanning more than a few months.

Also invoke unprompted when your own draft answer is about to name a specific product, credential, exam code, price, or market claim. The user asking a soft question does not make the underlying facts soft. This is the common miss: the skill fires on what you are about to assert, not only on what was asked.

“我应该考取哪种认证？”·“X是否仍受支持？”·“Y的市场需求如何？”·“我们是否应采用Z？”·“……领域的趋势是什么？”·“五年后我的职业发展方向应该是什么？”·任何跨度超过几个月的路线图。

即便未被提示也需调用：当你准备在草稿回答中提及特定产品、资质、考试代码、价格或市场主张时。用户提出的问题看似随意，并不代表背后的事实也可以随意对待。这是常见的遗漏点：本技能触发的依据是你即将做出的断言，而不仅仅是用户的提问。

When not to

何时不调用

The answer lives in the user's own code or system →
```
study-design-verify
```
A single fact lookup with no recommendation riding on it → just look it up
Pure preference or taste, no external facts at stake → just answer

答案存在于用户自身代码或系统中 → 使用
```
study-design-verify
```
技能
仅进行单一事实查询，未附带建议 → 直接查询即可
纯偏好或品味问题，不涉及外部事实 → 直接回答即可

Why this is hard

为什么这很困难

Two forces work against you, and they pull in the same direction:

Your knowledge has no timestamp. You remember a given exam code as the architect certification because it was, for years. Nothing in that memory announces that it retired. Recall feels identical whether the fact is current or three years dead.

Confident structure hides rot. A well-organised recommendation — tiers, timelines, trade-offs — reads as researched. The scaffolding lends credibility to whatever sits inside it, including a dead product at the centre.

So the discipline cannot be "be careful." It has to be a mechanical pass that runs before the advice ships.

有两种力量在阻碍你，而且它们的方向一致：

你的知识没有时间戳。你记得某个考试代码是架构师认证，因为多年来它一直是。这段记忆中没有任何迹象表明它已停用。无论事实是当前有效的还是已失效三年，回忆的感觉都是一样的。

自信的结构会掩盖过时信息。组织有序的建议——分层次、有时间线、权衡利弊——读起来像是经过调研的。这种框架会赋予其中内容可信度，即使核心是已停产的产品。

因此，不能只是“小心谨慎”，必须在建议输出前执行一套机械的检查流程。

The method

方法流程

Six stages. Skipping one is a decision; make it out loud.

六个阶段。跳过任何一个阶段都需明确做出决策。

1. Inventory the moving parts

1. 梳理动态要素

Before researching anything, list every external entity your recommendation will name: products, credentials, exam codes, SKUs, APIs, prices, programs, market claims. This list is your verification queue.

Do this first because it is much harder to notice an unverified fact once it is embedded in a persuasive paragraph. Extract them while they are still just a list.

在开展任何调研前，列出你的建议将提及的所有外部实体：产品、资质、考试代码、SKU、API、价格、计划、市场主张。这份清单就是你的验证队列。

先做这一步是因为，一旦未验证的事实嵌入到有说服力的段落中，就很难再察觉。趁它们还只是清单时提取出来。

2. Lifecycle-verify every entry

2. 对每个条目进行生命周期验证

For each item: does it still exist, in the form you remember, today?

Go to the vendor's retirement registry — the authoritative list of what is dead and what is scheduled — not a blog, not a summary, and not memory. Most vendors publish one:

Microsoft:
```
credentials/support/retired-certification-exams
```
and
```
credential-retirement
```
; individual exam study guides also carry a retirement warning banner at the top
Cloud/API vendors: deprecation schedules, API version support matrices, end-of-life pages
Partner/licensing programs: the program's own requirements page, which changes quietly

Why this stage is first and non-negotiable. In the session that produced this skill, the recommendation was about to be "target PL-600." PL-600 had retired six weeks earlier — along with roughly ten other certifications inside a single quarter. Nothing about recalling it felt uncertain. One registry lookup caught it; no amount of careful reasoning would have.

When a vendor rewrites a whole product or credential line at once, that is itself a finding worth reporting — it is the clearest statement of direction you will get for free.

针对每个条目：它是否仍以你记忆中的形式存在于当下？

前往供应商的停用登记处——权威的停用及待停用清单——而非博客、摘要或依赖记忆。大多数供应商都会发布此类清单：

Microsoft：
```
credentials/support/retired-certification-exams
```
和
```
credential-retirement
```
；单个考试学习指南顶部也会带有停用警告横幅
云/API供应商：弃用时间表、API版本支持矩阵、终止服务页面
合作伙伴/许可计划：计划自身的要求页面，该页面会悄然变更

为什么此阶段是首要且不可协商的。在催生本技能的会话中，原本准备给出的建议是“瞄准PL-600”。但PL-600早在六周前就已停用——同一季度内还有约10项其他认证也被停用。回忆这个考试代码时没有任何不确定感。一次登记处查询就发现了问题；再多的谨慎推理也无法做到。

当供应商一次性重构整个产品或资质系列时，这本身就是一个值得报告的发现——这是你能免费获取的最清晰的方向声明。

3. Hunt the counter-signal

3. 寻找反向信号

Identify whose signal your recommendation currently rests on, and go find someone with a different incentive.

A vendor's own materials tell you what the vendor wants to be true. That is genuinely informative — but it is one party, and usually the most optimistic one. Before advising, deliberately look for the contradicting view: independent analysts, adoption data, post-mortems, practitioner reports.

In the same session, the first five-year recommendation came entirely from one vendor's credential portfolio, read at the top of a hype wave. Deliberately hunting a counter-signal surfaced an analyst finding that the technology sat at peak expectations with over 40% of projects expected to be cancelled within eighteen months. That did not reverse the direction — but it changed the advice from "learn the shiny thing" to "learn to make the shiny thing survive production," which is a materially different plan.

If you cannot find a counter-signal, say so. "I looked for a contradicting view and did not find one" is a real finding. "I did not look" is not the same thing, and must not be reported as if it were.

确定你的建议当前所依赖的信号来源，然后去寻找持有不同动机的主体的观点。

供应商自身的材料会告诉你供应商希望成为现实的情况。这确实有参考价值——但这只是一方的观点，而且通常是最乐观的一方。在给出建议前，刻意寻找相反的观点：独立分析师报告、采用数据、事后复盘、从业者反馈。

在同一会话中，最初的五年期建议完全来自某供应商的资质组合，且正值炒作高峰期。刻意寻找反向信号后，发现分析师的研究结果显示该技术处于预期峰值，超过40%的项目预计在18个月内被取消。这并没有改变方向——但建议从“学习热门技术”变成了“学习如何让热门技术在生产环境中存活”，这是本质上不同的规划。

如果找不到反向信号，请如实说明。“我寻找了相反观点但未找到”是一个真实的发现。“我没有寻找”则完全不同，绝不能将其伪装成前者。

4. Count primary artifacts before citing commentary

4. 引用评论前先统计原始数据

For any claim about what a market wants, count the actual artifacts before quoting anyone's summary of them.

Job postings, changelogs, release notes, package download counts, repository activity, pricing pages. Fifty real postings tell you more than any number of trend pieces — and trend pieces are written to rank in search, not to be accurate.

This one reversed a published conclusion mid-session. Trend articles supported an "AI governance is in demand" positioning. Counting the actual listings — about 25 in one national market, 23 in another — found that phrasing in essentially none of them. The positioning was a forward bet on a later year, not a present-day hiring filter. Same underlying direction; completely different advice about timing.

Counting is usually cheap. Do it before, not after, someone asks whether you did.

对于任何关于市场需求的主张，先统计实际数据，再引用他人的总结。

招聘信息、更新日志、发布说明、包下载量、仓库活跃度、定价页面。50条真实的招聘信息比任何数量的趋势文章都更有价值——而趋势文章的撰写目的是提升搜索排名，而非保证准确性。

这一步在会话中推翻了已发布的结论。趋势文章支持“AI治理需求旺盛”的定位。统计实际招聘信息——某国家市场约25条，另一市场约23条——发现其中几乎没有使用该表述的。该定位是对未来几年的前瞻性押注，而非当前的招聘筛选标准。方向相同，但关于时机的建议完全不同。

统计通常成本很低。在有人询问你是否做过统计之前就完成它。

5. Read the institutional incentive, not just the person's

5. 解读机构动机，而非仅关注个人目标

The person asking has goals. The organisation around them has requirements, deadlines, and exposures — and those are often more actionable, because institutions pay for things individuals merely want.

Ask: what does the employer, the team, the customer, or the program need? What are they at risk of losing? What deadline is already running against them?

The single most useful output of the originating session came from this stage. Reading the employer's partner-program requirements revealed that three of the four credentials qualifying for a scoring category had just retired, creating a dated cliff. That turned "please fund my exam" into "I can protect our designation" — a completely different conversation, and one the individual's own goals would never have surfaced.

提问者有自己的目标。他们所在的组织有要求、截止日期和风险——这些往往更具可操作性，因为机构会为需求付费，而个人只是想要。

思考：雇主、团队、客户或计划需要什么？他们面临失去什么的风险？已经在倒计时的截止日期是什么？

催生本技能的会话中，最有用的输出就来自这一阶段。解读雇主的合作伙伴计划要求后发现，某评分类别下符合条件的四项资质中有三项刚刚停用，形成了一个过时危机。这将“请资助我的考试申请”转变为“我可以保护我们的资质认证”——这是完全不同的对话，而仅关注个人目标永远不会发现这一点。

6. Compute headline numbers; never sum rounded parts

6. 计算关键数据；切勿将四舍五入后的数值相加

Any number that carries the recommendation gets computed from source values, in a script, once.

Rounding each component and summing them compounds the error into exactly the figure people will quote back at you.

A headline figure in that session's spec read 72.6%. Recomputing from the underlying weights gave 72.8% — the earlier number had been assembled from per-item gains that were each already rounded. The conclusion survived; the number in the acceptance criterion did not.

任何支撑建议的数值都应从源值出发，通过脚本计算一次。

将每个组件四舍五入后再相加，会将误差放大成人们会反复引用的数值。

该会话规范中的一个关键数据显示为72.6%。根据底层权重重新计算后得到72.8%——之前的数值是由已四舍五入的分项收益汇总而来。结论得以保留，但验收标准中的数值被修正。

Grade every claim

为每个主张评级

The point of grading is not decoration. It tells the reader which parts of your advice they can act on directly and which they must re-check themselves — and it forces you to notice when a load-bearing claim is resting on a blog post.

Grade	Means
Verified-primary	You read the authoritative source yourself — vendor registry, official docs, the actual postings
Corroborated	Two or more independent sources agree, but you did not read the primary
Directional	Single source, or an aggregator, or content with an incentive to inflate. Usable for direction; never as an anchor
Unverified	Asserted from memory or inference. Say so plainly, or cut it

Salary aggregators are the canonical directional case: in the originating session the same job title was reported at both ~$124k and ~$176k. A number that disagrees with itself by 40% can indicate a trend, but it must never anchor a negotiation, and saying so is part of the deliverable.

Corroboration between independent sources is worth flagging explicitly. When two unrelated parties measuring different things land on the same answer, that agreement is stronger evidence than either alone — and readers will miss it unless you point it out.

评级的目的不是装饰。它告诉读者你的建议中哪些部分可以直接采纳，哪些部分必须自行重新核查——同时也迫使你注意到某个关键主张是否仅基于博客文章。

评级	含义
Verified-primary（已验证-原始来源）	你亲自查阅了权威来源——供应商登记处、官方文档、真实招聘信息
Corroborated（已佐证）	两个或多个独立来源达成一致，但你未查阅原始来源
Directional（方向性）	单一来源、聚合器或有夸大动机的内容。可用于判断方向；绝不能作为核心依据
Unverified（未验证）	基于记忆或推断的断言。请明确说明，或删除该内容

薪资聚合器是典型的方向性案例：在催生本技能的会话中，同一职位的薪资报告分别约为12.4万美元和17.6万美元。相差40%的数值可以指示趋势，但绝不能作为谈判的核心依据，这一点必须纳入交付成果。

独立来源之间的佐证值得明确标注。当两个无关主体通过不同方式得出相同结论时，这种一致性比任何单一来源的证据都更有力——除非你指出来，否则读者会忽略这一点。

Deliverable shape

交付成果的要求

Whatever the format, three things are required:

Every load-bearing claim carries its source and its grade. Inline is fine; a table is fine. Ungraded assertions read as verified whether or not they are.
Weak evidence is named, not smoothed. "Directional only — sources disagree by 40%" is more useful than a confident average that hides the spread.
State what you did not check. Geographies you skipped, sources that blocked you, questions you left open. A gap the reader knows about is a caveat; a gap they discover later is a defect.

Where a recommendation depends on something with a date attached — a retirement, a deadline, a cliff — surface the date prominently. That is the part that will silently invalidate the advice.

无论格式如何，必须包含三点：

每个关键主张都附带来源和评级。内联标注或表格形式均可。未评级的断言会被默认为已验证，无论实际情况如何。
明确标注弱证据，而非掩盖。“仅具有方向性——来源相差40%”比隐藏差异的自信平均值更有用。
说明未核查的内容。跳过的地区、无法访问的来源、未解答的问题。读者知晓的缺口是免责声明；读者事后发现的缺口则是缺陷。

当建议依赖带有日期的内容——停用日期、截止日期、危机点——请突出显示该日期。这是会悄无声息使建议失效的部分。

Red flags

警示信号

Recognise these as the moment to run the method, not skip it:

Thought	Reality
"I know this product well"	Knowing it well is exactly how you miss that it retired. Familiarity is not currency.
"The user only asked casually"	The facts are not casual. Soft question, hard facts.
"This is the obvious industry direction"	Obvious directions are usually one vendor's marketing, well distributed.
"The trend articles all agree"	Articles cite each other. Agreement among secondary sources is not corroboration.
"Counting postings is overkill"	It is usually ten minutes, and it has reversed conclusions.
"Close enough on the number"	The rounded number is the one that gets quoted back at you.
"I'll caveat it at the end"	A caveat under a confident recommendation does not survive being skimmed. Grade inline.

请将这些视为必须执行本方法的时刻，而非跳过的理由：

想法	现实
“我对这款产品非常了解”	正是因为了解深入，你才会忽略它已停用的事实。熟悉度不是可信度的保障。
“用户只是随口问问”	事实并非随口之事。问题看似随意，事实却至关重要。
“这是显而易见的行业方向”	显而易见的方向通常是某供应商广泛传播的营销内容。
“所有趋势文章都达成一致”	文章之间会互相引用。二手来源之间的一致并非佐证。
“统计招聘信息太小题大做了”	这通常只需要十分钟，却曾推翻过结论。
“数值差不多就行”	四舍五入后的数值会被人们反复引用。
“我会在最后加上免责声明”	自信建议下的免责声明在快速浏览时会被忽略。请内联标注评级。

Relationship to neighbouring skills

与相邻技能的关系

career-growth
— a five-station quarterly career review: evidence-graded skill inventory, job-market survey, a four-test "moat" selection gated on the user's own pick, then a certification-driven study plan. It owns the person-side analysis and the decision structure; it delegates all outside-world fact verification — certifications, exam codes, market demand — to this skill. Precedence: reach for
```
career-growth
```
for the full periodic review; reach for this skill for a single verified recommendation, or to check whether one named product or credential is still current.
study-design-verify
— same evidence-grounded stance, aimed at the user's own system. That one studies code, schemas, and live data. This one handles facts nobody in the room controls. They compose: study the system with that, verify the outside world with this.
scrutinize
— reviews a plan or change already on the table. This runs earlier, while the advice is still being formed.
reflect
— where the lessons from a mis-advised session get routed, including into this skill.

career-growth
—— 包含五个环节的季度职业回顾：基于证据评级的技能盘点、就业市场调研、基于用户自身选择的四项“护城河”测试，然后是基于认证的学习计划。该技能负责个人层面的分析和决策结构；所有外部事实验证——认证、考试代码、市场需求——均委托给本技能。优先级：如需完整的定期回顾，请使用
```
career-growth
```
；如需单一的已验证建议，或检查某一指定产品/资质是否仍有效，请使用本技能。
study-design-verify
—— 同样基于证据，针对用户自身系统。该技能研究代码、架构和实时数据。本技能处理无人可控的外部事实。二者可结合使用：用前者研究系统，用后者验证外部世界。
scrutinize
—— 审查已提出的计划或变更。本技能更早运行，在建议形成过程中执行。
reflect
—— 接收来自错误建议会话的经验教训，包括整合到本技能中。