Skillrush Town

Skillrush Town

消费侧用法（用户只想看榜时）

当用户问"今天有什么新 skill / 什么在涨 / 有什么值得装"这类问题时，不要走维护流程，直接读数据回答：

1. 取数据：仓库内优先读 data/latest.json；不在仓库内时读 https://learnprompt.github.io/skillrush-town/data/latest.json。 回看历史用 data/snapshots/<YYYY-MM-DD>.json，可用日期列表在 data/dates.json。
2. 总结三块：今日 Top10（items 前 10 条）、潜力 Skill（报告 data/reports/<date>.md 的「潜力 Skill」节，或按 SKILL.md 潜力标准从 items 现算）、 新进榜（有历史时 prev_rank 为 null 的条目）。
3. 给出"值得点开看"的 2-3 个，并说清理由（下载/星标增量、排名变化、summary 里的用途）， 附上 https://learnprompt.github.io/skillrush-town/?date=<snapshot_date> 方便用户自己看。
4. 诚实转述 limitations 与 comparison_basis.note：缺历史切片时不要把数字说成日环比。

First Reads

When this skill triggers inside the repo, read:

• README.md for product positioning and user paths.
• scripts/clawhub_daily.py before changing data generation.
• data/dates.json and the latest data/snapshots/*.json before changing the UI.
• assets/app.js, assets/styles.css, and index.html before changing the page.
• references/source-contract.md before changing ClawHub request semantics.
• references/source-adapter-pattern.md before adding changelog, model leaderboard, or any non-ClawHub monitoring source.
• references/publishing.md before changing GitHub Pages or Actions.

Product Direction

Maintain three surfaces:

• Public town board: a phone-friendly GitHub Pages site where anyone can read the latest ClawHub Top100, growth lists, potential Skills, and history.
• Forkable Skill generator: a repo plus Skill that lets users generate their own Skill radar without copying private tokens or personal notes.
• Source-monitoring workflow: a repeatable pattern for turning public leaderboards, changelogs, and release pages into dated snapshots, diffs, reports, and optional Agent reminders.

Keep the default story simple: "每天从公开榜单和更新日志里淘出值得看的 AI 变化." ClawHub Top100 is the first source because it is complex enough to prove the workflow: runtime request, Convex path, cursor pagination, Top100 normalization, historical diffs, reports, and a public board.

Do not position this as "just a webpage". The page is the town board; the Skill is the workflow contract that lets future agents maintain and extend the town.

Good future source examples:

• Claude Code changelog: monitor dated release notes and new capabilities.
• Artificial Analysis model leaderboard: monitor model ranking, price, speed, and benchmark changes.
• Other public marketplaces or leaderboards with stable sort and pagination.

First Run Usage

Installing this Skill only installs the workflow contract; it must not create hidden scheduled jobs during installation.

For a one-off check, the user can say:

textCopy code
Use skillrush-town to read latest.json and summarize today's ClawHub Top10 and potential Skills.

For a daily reminder in Hermes, create a cron job only when the user asks for it. A good default is 10:00 Asia/Shanghai, after the GitHub Actions update window:

textCopy code
Every day, read https://learnprompt.github.io/skillrush-town/data/latest.json,
summarize the snapshot date, Top10, potential Skills, limitations, and link to
https://learnprompt.github.io/skillrush-town/?date=<snapshot_date>.

Codex and Claude Code are usually task runners, not persistent reminder daemons. For them, keep the repo forkable and use GitHub Actions, system cron, or Hermes cron for notifications.

Source Rules

Default ClawHub Source

• The canonical ranking source is Convex api/query, path skills:listPublicPageV4.
• Request args must keep sort=downloads, dir=desc, nonSuspiciousOnly=true, highlightedOnly=false, numItems=25.
• Build Top100 by following nextCursor for 4 pages.
• Do not use GET /api/v1/skills as the primary ranking basis.
• If API fields, path, or pagination change, write the limitation into both snapshot and report.

Adding New Sources

Before adding a changelog, model leaderboard, release feed, or any non-ClawHub source, read references/source-adapter-pattern.md and create a source-specific contract file:

textCopy code
skills/skillrush-town/references/source-contract-<source>.md

Do not merge a new source adapter unless it has:

• a canonical URL or request contract
• stable item or entry keys
• snapshot schema
• diff semantics
• limitation handling
• headless tests with fixtures or mocked network responses

Daily Report Rules

Every run must produce:

• data/snapshots/YYYY-MM-DD.json
• data/reports/YYYY-MM-DD.md
• data/latest.json
• data/dates.json

Reports must include new entries, dropped entries, Top10 changes, downloads growth Top10, stars growth Top10, and potential Skills. If there is no potential Skill, explicitly write 今日无新增潜力skill.

Never describe the first run, migration run, or missing-history comparison as a strict daily delta.

Potential Skill Criteria

Include at most 10. A Skill qualifies if any condition is true:

• new Top100 entry
• download delta Top20 and star delta Top30
• rank rises by at least 8 places

Each potential Skill needs name, rank change, download/star delta, and one short recommendation.

Writing Style

For public README and release text, keep the town/gold-rush metaphor but avoid AI-marketing boilerplate. Prefer concrete use cases over abstract claims. Do not say "empower", "ecosystem flywheel", "comprehensive platform", or similar filler.

Validate

Required Headless Validation

These checks must work without Chrome, Playwright, Puppeteer, Camofox, Selenium, or browser login state:

bashCopy code
python -m py_compile scripts/clawhub_daily.py
python -m pytest -q
python "${CODEX_HOME:-$HOME/.codex}/skills/.system/skill-creator/scripts/quick_validate.py" skills/skillrush-town

For a live ingestion check, write into a temporary data directory instead of mutating committed data:

bashCopy code
TMP=$(mktemp -d)
python scripts/clawhub_daily.py --date 2026-05-04 --data-dir "$TMP/data"

Optional Browser / Manual Page Check

Only run this when a browser is available:

bashCopy code
python3 -m http.server 8093

Open http://127.0.0.1:8093/?date=2026-05-04 and verify the date selector, Top10, limitation panel, potential Skill section, search, and Top100 table.