ask-exemplar

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

ask-exemplar

Consult current external evidence before choosing or judging a Target. Every run is a Consultation: it is source-grounded and does not change the Target. A standalone run returns a Decision Brief; an Embedded Evaluation returns only Top Fixes.

在选择或评判Target之前，先参考当前的外部证据。每次运行都是一次咨询：基于可靠来源，且不会修改Target。独立运行会返回一份Decision Brief；嵌入式评估仅返回Top Fixes。

Workflow

工作流

Resolve the Target. Infer whether it is open or completed. Use Guidance for an open Target and Evaluation for a completed Target. Ask one question only when an unresolved ambiguity would change the research. Identify the domain, mandatory requirements, and the user's hard constraints. For Guidance, identify the options. For Evaluation, read the completed artifact or diff and its project context.
Research external evidence. Search or revalidate current online sources on every invocation. Prefer primary sources such as formal standards, official documentation, shipped work, direct measurements, and an Exemplar's published principles. Classify the Standard Finding as
```
formal requirement
```
,
```
dominant convention
```
,
```
contested practice
```
, or
```
no Standard
```
. Use
```
undetermined
```
only when an Evidence Gap prevents classification. Select one or two Exemplars for evidence, relevance, and transferability. A person, organization, product, or method can be an Exemplar, but reputation alone is not evidence. Record the supporting sources as the Verified Source Set.
Derive observable criteria. Convert the supported evidence into criteria that can be checked against the Target. Separate mandatory requirements, conventions, and Exemplar traits. Place a source link next to each material claim. For code, UI, documentation, API, CLI, scope, reliability, tests, short writing, or agent-skill Targets, read REFERENCE.md. Do not let any criterion depend on simulated advice from a famous person.
Apply the criteria. Remove options that violate mandatory requirements or the user's hard constraints, then use conventions and Exemplars to rank the Feasible Options. State the tradeoff and Confidence for each material finding. A fix that violates a mandatory requirement or a hard constraint is not a fix; name the tension instead.
- In Guidance, compare the options and select one by default. Make the Recommendation conditional only when one unresolved choice changes the preferred option.
- In Evaluation, check the completed Target against each criterion, give one concrete fix per gap, and rank the deduplicated Top Fixes by leverage. Also check whether a maintained community tool, library, pattern, or idiom can replace custom work in the Target; a supported replacement is a Top Fix candidate, and its dependency and maintenance cost is a tradeoff.
Return the Decision Brief. Use the format below in the conversation unless the user asks for a file. Under Embedded Evaluation, return only the Top Fixes instead.

确定Target类型。推断它是待决策的（open）还是已完成的（completed）。对于待决策的Target使用Guidance，对于已完成的Target使用Evaluation。只有当存在未解决的歧义且该歧义会影响研究方向时，才提出一个问题。明确领域、强制性要求以及用户的硬性约束。对于Guidance，确定可选方案；对于Evaluation，阅读已完成的工件或代码差异及其项目背景。
研究外部证据。每次调用时都要搜索或重新验证当前的在线来源。优先选择一手来源，如正式标准、官方文档、已发布的成果、直接测量数据以及范例（Exemplar）公开的原则。将标准发现（Standard Finding）分类为
```
formal requirement
```
、
```
dominant convention
```
、
```
contested practice
```
或
```
no Standard
```
。只有当存在证据缺口（Evidence Gap）导致无法分类时，才使用
```
undetermined
```
。选择1-2个范例，依据是证据充分性、相关性和可迁移性。个人、组织、产品或方法都可以作为范例，但仅凭声誉不能作为证据。将支持性来源记录为已验证来源集（Verified Source Set）。
推导可观测标准。将已验证的证据转化为可与Target对照检查的标准。区分强制性要求、惯例和范例特征。每个实质性主张旁都附上来源链接。如果Target是代码、UI、文档、API、CLI、范围、可靠性、测试、短篇写作或agent-skill，请阅读REFERENCE.md。任何标准都不得依赖模拟的名人建议。
应用标准。先排除违反强制性要求或用户硬性约束的选项，然后利用惯例和范例对可行选项（Feasible Options）进行排序。说明每个实质性发现的权衡和置信度（Confidence）。违反强制性要求或硬性约束的“修复方案”不能算作修复，应说明其中的矛盾。
- 在Guidance中，对比各选项并默认选择一个。仅当存在未解决的选择会改变首选方案时，才将建议设为有条件的。
- 在Evaluation中，对照每个标准检查已完成的Target，针对每个差距给出一个具体的修复方案，并按影响力对去重后的Top Fixes进行排序。同时检查是否有维护中的社区工具、库、模式或 idiom 可以替代Target中的自定义工作；受支持的替代方案是Top Fixes的候选，其依赖和维护成本是需要权衡的因素。
返回Decision Brief。除非用户要求文件，否则在对话中使用以下格式。在嵌入式评估中，仅返回Top Fixes。

Evidence rules

证据规则

Use credible secondary sources only as provisional context when primary evidence is unavailable. Label the Evidence Gap and lower Confidence.
When online access is unavailable, report the Evidence Gap and make no claim about the current Standard or strongest supported Exemplar. Use available local sources only as provisional context, set Confidence to
```
low
```
, and give the reason.
Call an Exemplar the best in the world only when the evidence proves that claim. Otherwise call it the strongest supported Exemplar for this Target and state the Coverage Limit.
Stop research when current evidence is sufficient to classify the Standard Finding and support one or two relevant Exemplars. State important limits instead of implying an exhaustive search.
Reuse a Verified Source Set only in the current conversation or parent workflow, and only when the domain and Standard Finding have not changed. Recheck time-sensitive claims before reuse.
Use
```
high
```
,
```
medium
```
, or
```
low
```
Confidence and give a short reason. Do not use a numeric score for Confidence.

只有当一手证据不可用时，才将可信的二手来源作为临时背景使用。标注证据缺口（Evidence Gap）并降低置信度（Confidence）。
当无法访问网络时，报告证据缺口，不对当前标准或最具认可度的范例做出任何声明。仅将可用的本地来源作为临时背景使用，将置信度设为
```
low
```
并说明原因。
只有当证据能证明某范例是全球最佳时，才称其为全球最佳；否则，称其为针对该Target最具认可度的范例，并说明覆盖范围限制（Coverage Limit）。
当当前证据足以对标准发现进行分类并支持1-2个相关范例时，停止研究。说明重要限制，而非暗示已进行全面搜索。
仅在当前对话或父工作流中复用已验证来源集，且仅当领域和标准发现未发生变化时才可复用。复用前需重新检查时效性强的主张。
使用
```
high
```
、
```
medium
```
或
```
low
```
表示置信度，并给出简短理由。不要用数字评分表示置信度。

Decision Brief

markdown

undefined

markdown

undefined

Decision Brief

Standard Finding

<classification, evidence, and Confidence>

<分类、证据和置信度>

Exemplars

<1-2个选择、适配原因、证据和置信度>

Observable criteria

<标准以及与Target的对比>

Recommendation

<首选可行选项、权衡和置信度>

Evidence limits


For Evaluation, make the criteria section a table with `criterion`, `bar`, `current`,
and `fix` columns, state the highest-leverage change in the Recommendation section,
and close the brief with a `### Top Fixes` section. Each Top Fixes item contains the
finding, concrete fix, Confidence, and tradeoff. REFERENCE.md shows a filled example.

<证据缺口和覆盖范围限制，或“无实质性限制”>


对于Evaluation，将标准部分设为包含`criterion`、`bar`、`current`和`fix`列的表格，在Recommendation部分说明影响力最高的变更，并在简报末尾添加`### Top Fixes`部分。每个Top Fixes条目包含发现、具体修复方案、置信度和权衡。REFERENCE.md展示了一个填充好的示例。

Embedded Evaluation

嵌入式评估

When a parent workflow requests Embedded Evaluation, remain report-only and return only the numbered Top Fixes list. Each item contains the finding, concrete fix, Confidence, and tradeoff. Do not edit files, commit, or push; the parent workflow controls all changes. When there are no material fixes, return exactly

No Top Fixes.

当父工作流请求嵌入式评估时，仅返回报告内容，即编号的Top Fixes列表。每个条目包含发现、具体修复方案、置信度和权衡。不要编辑文件、提交或推送；所有变更由父工作流控制。当没有实质性修复方案时，返回确切内容

No Top Fixes.

Self-check

自我检查

An Evaluation is useful only when it surfaces what correctness, simplification, and maintainability reviewers structurally cannot: a missing standard tool, pattern, or an unmet external bar. A Guidance run is useful only when its criteria come from cited evidence instead of restating what the user already knows. When a run fails this test, return to step 2 and research the domain's primary sources, or state that no external bar applies.

只有当评估能发现评审人员在结构上无法发现的问题（如缺少标准工具、模式或未达到外部要求）时，该评估才有用。只有当Guidance的标准来自引用的证据而非重复用户已知信息时，该Guidance运行才有用。如果某次运行未通过此测试，请回到步骤2研究该领域的一手来源，或说明不存在外部要求。