Multi-platform creator analytics → one self-contained, presentable HTML report page.
多平台自媒体数据复盘 → 一张自包含、可公开展示的 HTML 报告页。
creator-analytics is a Claude Code / Agent skill. Once your cross-platform numbers are crunched, it locks down the last mile — turning scattered conversation output into a single static HTML page with a clear point of view: a finding-card lede, a heat-mapped master table, topic-by-topic charts, a per-platform iteration panel, and a tickable execution checklist. Pure Python standard library generates the HTML, charts are inline SVG, and zero JS libraries are loaded — open the file in any browser.
creator-analytics 是一个 Claude Code / Agent 技能。当你把跨平台数据算完之后,它把最后一公里固化下来 —— 把散落的对话输出收成一张静态 HTML 页,有明确判断:核心结论卡先行、热力总表、专题图表、分平台迭代面板、可勾掉的执行 Checklist。纯 Python 标准库生成 HTML,图表用内联 SVG,零 JS 库 —— 任何浏览器打开即可。
The hard part of a data review isn't generating the page — it's making sure the insight survives the presentation layer, and that the disciplined steps don't get skipped. Most analyses end as a wall of chat messages and a few stray screenshots; the judgment, the calibrations, and the actionable next steps evaporate the moment the session ends. creator-analytics takes the opposite stance: the deliverable is a real, reviewable artifact — one HTML page that leads with judgment, colors every magnitude column on a log scale so one viral hit doesn't whitewash the rest, flags anomalies honestly, and ends with a checklist you can actually tick off.
数据复盘真正难的不是把页面画出来 —— 是让洞察别在呈现层掉价,以及那些有纪律的步骤别被跳过。多数分析最后只剩一墙对话和几张零散截图;判断、口径校正、可落地的下一步,会话一结束就蒸发了。creator-analytics 反其道而行:交付物是一个真正可复审的成品 —— 一张 HTML 页,开篇就给判断、每一列量级指标按对数刻度上色(一个爆款不会把整列洗白)、异常如实标注,最后是一份你真能勾掉的清单。
- Insight methodology baked in — the
references/insight-playbook.md0→9 analysis path: read magnitude and rate separately, locate the funnel stage before judging content, use natural-experiment control groups to nail causation, check concentration (75% of new fans from one post ≠ a stable engine), compute correlations on your own data to test platform folklore, and attribute findings down to a changeable action. 内置洞察方法论 ——references/insight-playbook.md的 0→9 分析路径:量级与率分开读、先定位漏斗层级再评判内容、用自然实验对照组坐实因果、查集中度(75% 涨粉来自一篇 ≠ 稳定引擎)、用自己的数据现算相关性验平台传言、把结论归因到可改的动作。 - A presentable page skeleton — finding cards (judgment first) → master heat table (many entities × many metrics, colored per-column on a log scale) → topic sections (bar chart / SVG scatter + threshold line) → per-platform iteration panel (P0 in red) → execution checklist → footer cadence statement. 可展示的页面骨架 —— 核心结论卡(先给判断)→ 热力总表(多实体 × 多指标,列内对数刻度上色)→ 专题分析(条形图 / SVG 散点 + 阈值线)→ 分平台迭代建议(P0 标红)→ 执行 Checklist → footer 口径声明。
- The authoritative PPVI light design system —
assets/template.htmlis the single style source of truth: zero dependencies, no JS library, opens straight in a browser. This repo is itself the origin of the PPVI light style, so its own pages are meant to be exemplary. 权威 PPVI 浅色设计系统 ——assets/template.html是唯一样式权威源:零依赖、无 JS 库、浏览器直接打开。本仓库就是 PPVI 浅色样式的源头,它自己的页面应当是范例级。 - Copy-ready component helpers — heat coloring, bar chart, scatter plot, four-semantic badges, iteration panel, and checklist renderer; the generation script copies them straight from
references/components.md. 可直接抄的组件 helper —— 热力上色、条形图、散点图、四语义徽标、迭代面板、Checklist 渲染器,生成脚本直接抄references/components.md。 - Analysis discipline as guardrails — single source of numbers (read only the parsed JSON, never hand-type), anomalies with no visible cause marked 待核实 (to verify) rather than guessed at, research conclusions tagged and kept apart from your own measured data, and iteration advice re-researched every cycle instead of reusing the old. 纪律即护栏 —— 数字单一来源(只读解析后的 JSON,不手填)、无可见原因的异常标「待核实」不臆断、调研结论加标注并与实测分开、迭代建议每周期重新调研不沿用存量。
What it locks in · 它固化了什么 — Reach for creator-analytics when the analysis is done and the conclusions need to land as a presentable, reusable HTML page — the report-page step of a cross-platform review. It fixes the things that should never be improvised: the page structure, the design tokens, the chart helpers, and the review discipline (judgment-first cards, honest anomaly handling, footer cadence statement). 当分析做完、结论需要落成一张可展示可复用的 HTML 页时就用它 —— 即跨平台复盘的报告页步骤。它固定那些不该即兴发挥的东西:页面结构、设计 token、图表 helper、审查纪律(结论先行卡、诚实的异常处理、footer 口径声明)。
What it does not do for you · 它不替你做什么 — The depth still comes from each cycle's real work: reading the data, computing correlations, doing attribution, researching the latest platform playbooks. It is not the blog-article page, not the publish copy, and not the raw data parsing (an upstream script handles that). It is not a generic charting library either. 分析的深度仍然来自每个周期的真实劳动:读数据、算相关性、做归因、调研平台最新打法。它不是博客文章页、不是发布文案、也不是原始数据解析(上游脚本的事),更不是一个通用图表库。
Clone this repository into your Claude Code skills directory — project-level (recommended) or user-level:
把本仓库克隆到你的 Claude Code 技能目录 —— 项目级(推荐)或用户级:
# Project-level (recommended) · 项目级(推荐)
git clone https://github.com/xntj-ai/creator-analytics .claude/skills/creator-analytics
# User-level, available to every project · 用户级,所有项目可用
git clone https://github.com/xntj-ai/creator-analytics ~/.claude/skills/creator-analyticsThat is the only setup. The generated report is pure static HTML — no runtime dependency, no CDN, nothing installed locally. The one tool the generation script uses is Python (3.8+, standard library only); a PEP 723 header lets you run it with uv run --script.
只需这一步。生成的报告是纯静态 HTML —— 无运行时依赖、无 CDN、本地不装任何东西。生成脚本唯一用到的工具是 Python(3.8+,仅标准库);加个 PEP 723 头就能用 uv run --script 跑。
With Claude · 对 Claude 说 — first parse each platform's backstage export into one aligned JSON (one piece of content = one entity, metrics grouped by platform), then say "build a review report page from this data." Claude follows the page structure and the hard rules in SKILL.md, generates a build_<topic>_html.py script, and produces the static HTML you open locally.
先把各平台后台导出解析成一份对齐的 JSON(一期内容一个实体,按平台分组指标),然后说"基于这份数据出一张复盘报告页"。Claude 会按 SKILL.md 的页面结构和铁律,生成一个 build_<topic>_html.py 脚本,产出你本地打开的静态 HTML。
# After Claude writes the script · Claude 写好脚本后
uv run --script scripts/build_<topic>_html.py
# → out: <topic>_review.html (open in any browser · 浏览器直接打开)Standalone, without Claude · 独立使用,不经 Claude — the design system lives entirely inside assets/template.html, so a model is optional:
设计系统完全封装在 assets/template.html 里,大模型可有可无:
- Copy
assets/template.html. · 复制assets/template.html。 - Fill the sections with your own findings and numbers (helper code in
references/components.md). · 用你自己的结论和数字填进各区块(helper 代码见references/components.md)。 - Open it in a browser. · 浏览器打开即可。
The page is assembled in a fixed order; drop any section whose data you don't have, but the finding cards and the footer cadence statement stay in every version.
页面按固定顺序组装;没有对应数据的区块可删,但核心结论卡和 footer 口径声明任何版本都保留。
| Section 区块 | Component 组件 | Purpose 作用 |
|---|---|---|
| Kicker + H1 + lead | — | Account / range / north-star question. 账号 / 区间 / 北极星问题。 |
| Stat strip | .stats |
3–4 global summary numbers. 3–4 个全局汇总数。 |
| Finding cards | .card (gold / j jade / t terra) |
4–6 judgment-first cards — the whole page in a glance. 4–6 张结论先行卡 —— 全页判断的浓缩。 |
| Master heat table | .tablecard |
Many entities × many metrics, per-column log-scale coloring + anomaly badges. 多实体 × 多指标,列内对数刻度上色 + 异常徽标。 |
| Topic sections × N | .bars / .scatter + .two |
One-line stance + finding-first lede + chart + paired panels. 一句话立场 + 结论先行 + 图 + 成对解读。 |
| Iteration panel | .iter (⚑→P0) |
Per-platform multi-dimension advice, P0 in red. 分平台多维建议,P0 标红。 |
| ★ Strategy cards | .card |
The cycle's key corrections (calibrations). 本轮校正后的关键修正。 |
| Execution checklist | .ckl |
Tickable, verb-first actions (P0 first). 可勾掉的动作清单(动词开头,P0 优先)。 |
| Footer | footer |
Data source + review-discipline statement + generation script. 数据来源 + 审查纪律声明 + 生成脚本名。 |
Color semantics run through the whole page: gold #C8922E neutral emphasis · jade #1f6f5c positive / confirmed · terra #B4502E warning / correction / P0.
色彩语义贯穿全页:gold #C8922E 中性强调 · jade #1f6f5c 正面/坐实 · terra #B4502E 警示/校正/P0。
The skill was distilled from a real 32-episode × 5-platform review on the「张拼拼·XNTJ」account — the same review is written up in the EP0034 article, where Claude Code acts as the operations director, merges every platform's backstage export, and produces the optimization checklist. The skill's own page (docs/) is built on this very assets/template.html, so it doubles as a live style reference.
本技能蒸馏自「张拼拼·XNTJ」账号一次真实的 32 期 × 5 平台复盘 —— 同一次复盘写在 EP0034 那篇文章里:Claude Code 当运营总监,合并各平台后台导出,产出优化清单。技能自己的页面(docs/)就是用这套 assets/template.html 搭的,可当作活的样式参考。
The generation script is plain Python standard library: it reads one aligned _<topic>_data.json, formats numbers, colors each magnitude column on a log scale (so a long-tail viral hit doesn't flatten the column to white), draws bar charts and scatter plots as inline SVG strings, and renders iteration panels and checklists from small data-driven helpers — then drops it all into the assets/template.html skeleton. The output is one static .html: no bundler, no backend, no CDN, no JS chart library. When the CSS goes into an f-string, the braces are doubled ({{); that is the only build wrinkle.
生成脚本就是纯 Python 标准库:读一份对齐的 _<topic>_data.json,格式化数字,把每一列量级指标按对数刻度上色(长尾爆款不会把整列压成白),用内联 SVG 字符串画条形图和散点图,用小巧的数据驱动 helper 渲染迭代面板和 Checklist —— 再全部填进 assets/template.html 骨架。产物是一个静态 .html:无打包、无后端、无 CDN、无 JS 图表库。CSS 放进 f-string 时花括号写成 {{,这是唯一的构建小坑。
Does the report need an internet connection? No. Unlike many "HTML report" tools, the output is pure static HTML with inline SVG charts and zero CDN calls — it opens offline and renders identically.
报告需要联网吗? 不需要。和很多"HTML 报告"工具不同,产物是纯静态 HTML,图表是内联 SVG,零 CDN 调用 —— 离线打开,渲染一致。
Can I use it without Claude? Yes. Copy assets/template.html, fill the sections with your own findings, and open it. The model is a convenience for turning analysis into a built page, not a dependency.
不用 Claude 能用吗? 可以。复制 assets/template.html,把各区块换成你自己的结论,打开即可。大模型只是帮你把分析变成成品页的便利,不是必需品。
Does the skill parse my raw exports? No — that's the upstream step. The hard rule is "single source of numbers": you parse each platform's export into one aligned JSON first, and the generation script only reads that JSON and computes, never hand-types a number. The footer states this explicitly.
技能会解析我的原始导出吗? 不会 —— 那是上游步骤。铁律是"数字单一来源":你先把各平台导出解析成一份对齐 JSON,生成脚本只读这份 JSON 算数,绝不手填任何数字。footer 会明确声明这一点。
Why log-scale coloring on the magnitude columns? Play counts and impressions are long-tailed; on a linear scale one viral entry pushes every other cell to near-white. Per-column log coloring keeps the within-column contrast readable. Magnitude metrics are colored per-column only — they are not comparable across platforms.
为什么量级列用对数刻度上色? 播放和曝光是长尾分布;线性刻度下一个爆款会把其它单元格压到接近白色。列内对数上色让列内对比依然可读。量级指标只做列内对照 —— 跨平台不可横比。
Public or private report? The page defaults to "will be shown publicly": monetization strategy (hooks / funnels / pricing) stays out of the public page; emit a separate private file when you need it.
报告公开还是私有? 页面默认按"会公开展示"对待:变现策略(承接钩子/漏斗/定价)不进公开页;需要时另出私有文件。
- ppvi — the PinPin Visual Identity methodology; creator-analytics is one of its light-mode origins. PinPin 视觉识别方法论,creator-analytics 是其浅色样式的源头之一。
- flowmaker — a sibling Claude Code skill that turns a process into a single-HTML editable flowchart, on the same light glass aesthetic. 同体系的 Claude Code 技能:把流程变成单 HTML 可编辑流程图,同款浅色玻璃风。
- cross-review — multi-model adversarial review; pairs well when a strategy call needs a second opinion. 多模型交叉审查,做策略决策时拉来交叉验证。
- xntj.tv — more Claude Code workflows and skills from 张拼拼 · XNTJ. 更多来自张拼拼·XNTJ 的 Claude Code 工作流与技能。