Content Flywheel · 内容飞轮

A local-first skill that turns every video into a calibrated experiment: score it, predict it blind, ship it, then let real data correct your taste. The compounding asset is your judgment, not any single post.

Live site · Skill · Rubric · Deploy · 中文 · Credits

Content Flywheel is a tiny, paper-thin workflow for solo creators. It does not ingest your analytics, run a server, or show you a dashboard. It does one thing: it forces you to commit a prediction before you see the numbers, then holds you to it. Over a few dozen posts, that single discipline turns vague gut feel into a scoring rubric that is measurably calibrated to your audience.

It is format-agnostic (B站, YouTube, 抖音, X, 小红书) and lives entirely in plain Markdown, so it diffs, backs up, and pairs with whatever production pipeline you already use.

Why this exists

Most "content tools" optimize the wrong loop. They give you a prettier view of what already happened. The bottleneck for a creator is not visibility, it is calibration: the gap between what you think will perform and what does. You only close that gap by predicting first and grading yourself honestly.

This repo is the smallest thing that enforces that loop end to end.

The loop

pick a topic        →  candidates.md
  │
make it             →  your own production pipeline
  │
score it            →  rubric.md (before publishing)
  │
predict it blind    →  prediction.md, then LOCK it (read-only)
  │
ship it             →  upload, record the link
  │  … wait 3 days …
retro at T+3        →  retro.md, real numbers vs your locked prediction
  │
evolve the rubric   →  only if the new rubric beats history (the gate)
  └────────────────────────────────────────────────►  next one lands closer

Three laws

1. Blind prediction. The forecast is written before any real data exists, and is frozen once locked. Hindsight is the enemy of calibration.
2. Upgrade equals full rescore. Changing a rubric rule means re-scoring every past sample with real data under the new rule. Adopt it only if its ranking matches actual outcomes on at least 4 of 5 samples. Otherwise reject it.
3. The rubric is a workbench, not an archive. Observations disproved by data are deleted, not hoarded. Git history is the archive.

Law 2 is the heart of it: it is the same validation gate that Microsoft SkillOpt uses to accept edits to a skill document, applied here to your content rubric. A rule is only "better" if it would have predicted the past better.

Install

As a skill (Claude Code / Codex / the skills ecosystem):

npx skills add https://github.com/appleweiping/content-flywheel

Or just clone and copy SKILL.md + skill/ into your project. There is nothing to build and no service to run.

git clone https://github.com/appleweiping/content-flywheel

Then, in your agent: initialize, learn from <account>, score this <path>, predict, shipped, retro, evolve, status. See SKILL.md for the full routing table.

Repository layout

content-flywheel/
├── SKILL.md                 installable skill manifest + routing
├── skill/
│   ├── rubric.md            the living, account-specific scoring rubric
│   ├── benchmark.md         competitor / personal-best anchors
│   ├── candidates.md        lightweight topic pool
│   └── templates/
│       ├── prediction-template.md
│       └── retro-template.md
├── examples/
│   └── three-body-bilibili.md   a worked specialization (AI-anime on B站)
├── web/                     static landing page (deploys to Vercel, no build)
├── assets/banner.png
├── DEPLOY.md
└── LICENSE

中文

内容飞轮是给个人创作者的极简工作流。它不抓你的后台数据、不跑服务、不给你看仪表盘。 它只做一件事：逼你在看到数据之前先写下预测，然后对账。几十条之后，这个纪律会把 "凭感觉"变成一把对你自己受众校准过的评分尺。

三铁律：①盲预测（发布前封板、只读）；②升级=全量重算（新尺必须在历史样本上 ≥4/5 跑赢旧尺才上线）； ③评分尺是工作台不是档案馆（被证伪的观察直接删，归档交给 git）。

平台无关（B站 / YouTube / 抖音 / X / 小红书），纯 Markdown。 安装：npx skills add https://github.com/appleweiping/content-flywheel。

Credits & lineage

This is a lightweight, honest companion to work that came first. If the heavier version fits you better, use theirs.

• XBuilderLAB/cheat-on-content — the score / blind-predict / retro / evolve loop. The full version adds command routing and hook-protected immutability.
• microsoft/SkillOpt — the validation gate: accept a change only if it beats held-out history.
• Leonxlnx/taste-skill — the anti-slop design rules this project's own landing page follows.
• abus-aikorea/voice-pro — production-side companion for voice cloning, captions, and multilingual dubbing.

License

MIT. See LICENSE. © 2026 Weiping Yan (appleweiping).