免费在 GitHub 公开

report-spec

Turn a request for a report or a dashboard widget into something buildable: the question it answers, the metric and its definition, where the data comes from, the query shape, and what it will not answer. Use when a new report or dashboard is requested, when an existing one needs a metric added or changed, when a widget is asked for in a sentence and needs pinning down, or when two reports disagree about the same number.

中文简介正在准备中。当前显示原文说明。

nanocoai/nanoclaw-templates

查看安装方法

包含的文件（1）

SKILL.md2.6 KB

SKILL.md（原文）

安装前可以查看交给 AI 代理的指令内容。

Report spec

Goal

A spec somebody can build against without a follow up conversation, and a number that means the same thing next quarter as it does today.

Procedure

Get the question behind the request. A request for a chart of complaints per site is really a question about which sites need attention, and the answer might not be that chart.
Define the metric precisely: the numerator, the denominator, the window, the grain and the filters. Check memory/conventions/metrics.md first, since the same word usually already means something here.
Say where each field comes from, which table or endpoint, at what grain, refreshed how often. A metric assembled from two sources with different refresh cadences will disagree with itself.
Check the grain before designing the query. If the join can produce more rows than entities, every total downstream is inflated, and this is the defect that survives review because the numbers look plausible.
Decide where it is computed. In the pipeline, in the query, or in the reporting layer. Computing the same metric in two places is how two reports end up disagreeing.
Say what it will not answer, so nobody reads a cause into a count.
Note the cost: how heavy the query is, how often it runs, and whether it needs its own pipeline step rather than running live.

Boundaries

Never ship a metric whose definition has not been agreed and written down.
Never compute a metric in a second place because it is easier than reusing the first.
Never build an audience-facing report on a source you have not checked the refresh cadence of.
Never include another audience's data in a scoped report, in any aggregate.
Do not build it as part of specifying it. The spec is the deliverable here.

What to record

Write every agreed definition into memory/conventions/metrics.md with its numerator, denominator, window, source and the date it was agreed, and note where it is computed. When two reports are found disagreeing, record the cause there too, because the same pair will disagree again after the next change.

评价

还没有评价。欢迎分享你的使用感受。

无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

同一仓库的技能

architecture

免费

Plan or review software architecture through a small expert panel. Use when the user explicitly asks for architecture planning, system design, technical strategy, an architecture assessment, a design review, or a system health review. Do not use for ordinary PR review, debugging, or implementation.

中文简介正在准备中。当前显示原文说明。