Deep learning pipeline

是什么

把"深度学习系列"每一期专题报告（访谈、论文、框架、争议）从内容生成、签名视觉风格、独立 CF Pages（Cloudflare Pages）子域名、4K + 移动 + 微信三档截图、到 Google Drive 结构化归档串成一条流水线，帮你把"一个深度学习选题"稳定产出成"一篇有编辑级排版、有独立访问入口、有可分发素材、有可追溯档案"的对外资产。

怎么用

1. 先按"内容是否自带视觉符号"决策：自带就委托一套新的签名风格（Hermes、Bloomberg、Anthropic 等），不带就从 html-style-router（风格路由）里选 12 套之一，让每期视觉不重样。
2. 用 deep-learning-skill 生成正文 + 3 层结构 + 引语 + 数据，再用所选风格统一渲染成 HTML，让正文质量和视觉档次同步保底。
3. 用 deploy-cf.sh 把这一期独立部署到 deeplearn-NN-slug.pages.dev（不与往期共域），让每期 URL 是干净的独立入口。
4. 用 render-poster.mjs 一次性产出 4K 主图、移动主图、1080w JPEG 微信版三档海报，让公众号、朋友圈、企业微信、AirDrop 各取所需。
5. 用 archive-gdrive.sh 按 "日期_深度学习_第NN期_标题" 命名归档到 Google Drive 深度学习目录，并在 state/memory 记一条 episode 索引，让后续每一期都能回溯。

架构图

flowchart LR
    A[选题] --> B[风格选型]
    B --> C[正文渲染 HTML]
    C --> D[CF Pages 独立部署]
    D --> E[三档截图]
    E --> F[GDrive 归档]

Deep Learning Pipeline

Single source of truth for the 深度学习系列 editorial pipeline. Each episode is a topic deep-dive (interview / paper / framework / public debate) shipped with a signature visual style (never reused across episodes), a dedicated CF Pages subdomain, and a triple-format poster (4K master + mobile master + WeChat-friendly JPEG).

Established episodes:

• #01 (2026-04-25) 罗福莉访谈 · DeepSeek 范式巨变 · OpenClaw 与 Agent 时代 — luofuli-interview-2026.pages.dev (legacy, pre-convention)
• #02 (2026-04-30) Hermes-Agent · Nous Research · OpenClaw 对比 · EvoMap 抄袭争议 — deeplearn-02-hermes-agent.pages.dev (Hermes Terminal style, frozen v11)

When to invoke

• User says "出第三期 / 下一期 / 新一期" (next episode signal)
• User refers to existing episodes by name (罗福莉式 / hermes-agent式) → continue the series with same delivery contract, fresh style
• User asks to publish a deep-learning topic article and expects CF deploy + GDrive archive (the full delivery loop, not just a doc)
• User says "深度学习管线 / learning pipeline / 专题报告 / 系列文章"

NOT for:

• Pure content generation without delivery → use deep-learning-skill directly
• One-off articles outside the editorial series → use html-style-router + bespoke pipeline
• Daily AI Intel / news digests → use wechat-content-pipeline or Hermes daily report

Why this skill exists (constraints the public web lacks)

The public web teaches "write Markdown → push to Pages". This skill encodes the non-obvious project constraints discovered through episodes #01 and #02:

1. Signature-style-per-episode is load-bearing, not decorative. If episodes share a style, the URL/file collision the user saw on #02 (luofuli-interview-2026.pages.dev/hermes-agent/) recurs every time. Each episode must commission OR pick a unique visual identity.
2. Sub-path bundling is a banned pattern. CF Pages projects must be one-per-episode, named by structured slug <cat>-<NN>-<title>.
3. background-attachment: fixed is forbidden on any custom style — it breaks fullPage screenshot capture (Chrome compositor bug). Encoded as v11 invariant in html-hermes-style/SPEC.md.
4. WeChat image-mode delivery requires 1080w JPEG, not 4K PNG. Three-layer rule (format / dimensions / send method) — encoded into the renderer script.
5. Wrangler env-var precedence trap: CLOUDFLARE_API_TOKEN env var with insufficient scope blocks deploys; OAuth fallback in ~/.wrangler/config/default.toml works if env vars are unset for the call.

Pipeline (5 stages, sequential)

A. STYLE      → pick stock or commission custom (one per episode, never reuse)
B. RENDER     → write HTML using picked style; auto-open
C. DEPLOY     → CF Pages dedicated project deeplearn-<NN>-<slug>; --branch=main
D. SCREENSHOT → single Playwright session: 4K + mobile + WeChat JPEG
E. ARCHIVE    → rclone to GDrive deep-learning folder + ledger entry

A. STYLE — pick or commission visual identity

Decision tree (references/style-decision.md):

• Source has signature visual identity (e.g. Hermes Greek mythology, BloombergTerminal-native data product, Anthropic warm-academic) → commission new bespoke skills/shared/html-<topic>-style/ with SPEC.md. Cost: ~2 hours; lifetime: reusable for any future content of that topic.
• No signature identity → pick stock from html-style-router (12 tier styles in 09-output-format.md). Match tone:
  • Interview / editorial → Claude Warm Academic OR The Economist
  • Strategy / market analysis → McKinsey Blue OR BCG Green
  • Pitch / fundraising → Sequoia Green-Gold
  • Data-dense / quantitative → Bloomberg Terminal
  • Technical / framework launch → Stripe Minimalist
• Banned: reusing the previous episode's style (the program owner rule "每次都是一个特色风格"). Episode N must NOT use episode N-1's style.

Output of stage A: chosen style ID + SPEC.md path. Logged in references/episodes.json.

B. RENDER — write HTML using picked style

Use the existing deep-learning-skill for content generation (3-layer structure / pull-quotes / stats / caveats), but override its template with stage A's style. Render path:

outputs/reports/<topic-slug>/<date>-<topic-slug>.html

Mandatory body invariants (any new style MUST honor):

• <html> { background-color: <bg-deep-token>; } — solid fallback for fullPage screenshot
• body uses background-color solid base + background-image gradients with background-size: 100% 100vh; no background-attachment: fixed (v11 invariant)
• Hero/header readable content and article body must share one layout frame token (.html-frame / .report-frame or style-equivalent); do not mix viewport hero padding such as 8vw with a separately centered main max-width
• text-rendering: optimizeLegibility on body
• Footer the program owner signature (no "Author:" prefix), per 09-output-format.md
• 2份制 sibling: .md canonical English version next to .html if the article is meant for cross-tool re-ingestion

After write: auto-open for visual review.

C. DEPLOY — CF Pages dedicated project

bash scripts/deploy-cf.sh <NN> <topic-slug> outputs/reports/<topic-slug>/

The script:

1. Stages the HTML as index.html in a temp folder
2. Creates project deeplearn-<NN>-<topic-slug> if it doesn't exist (production_branch=main)
3. Deploys with --branch=main (apex updates, not preview-only)
4. Verifies HTTP 200 + page title via curl
5. Auto-opens https://deeplearn-<NN>-<topic-slug>.pages.dev/

Auth handling: the script auto-unsets CLOUDFLARE_API_TOKEN / CF_API_TOKEN / CLOUDFLARE_ACCOUNT_ID for the call so wrangler falls back to OAuth in ~/.wrangler/config/default.toml. This bypasses the env-var precedence trap that blocked initial deploys on episode #02.

D. SCREENSHOT — triple-format poster

node scripts/render-poster.mjs --html outputs/reports/<topic-slug>/<date>-<topic-slug>.html

Single Playwright session producing three siblings:

File	Viewport	DSF	Type	Quality	Use
<date>-<topic-slug>-4k.png	1440×900	3	png	—	Master archive, future re-render reference
<date>-<topic-slug>-mobile.png	430×932 (isMobile)	3	png	—	AirDrop / file-mode WeChat
<date>-<topic-slug>-wechat.jpg	1080×1920 (isMobile)	1	jpeg	92	Image-mode WeChat / 企业微信 (1080w native, ≤2 MB, ratio ≤1:8)

All three written to the same outputs/reports/<topic-slug>/ folder.

E. ARCHIVE — GDrive structured rename + ledger

bash scripts/archive-gdrive.sh <NN> <topic-slug> outputs/reports/<topic-slug>/

Naming convention (mandatory, per the program owner 2026-04-30):

<YYYY-MM-DD>_深度学习_第<NN>期_<title-zh-or-mixed>.<ext>

Examples:

• 2026-04-30_深度学习_第02期_Hermes-Agent_Nous-Research_OpenClaw对比_EvoMap抄袭争议.html
• 2026-04-30_深度学习_第02期_Hermes-Agent_4k_screenshot.png
• 2026-04-30_深度学习_第02期_Hermes-Agent_wechat_screenshot.jpg

Target folder: gdrive: with --drive-root-folder-id=18Mnv4HyS4sGXOz7i7EiYZ_YATIifiBEv (the canonical deep-learning folder).

After archive, append a [DEEP-LEARNING-EPISODE] line to state/memory/<date>.md with episode number, title, style ID, CF URL, GDrive folder, and 4 file checksums.

Gotchas (built from real failures on episodes #01–#02)

1. background-attachment: fixed breaks fullPage capture. Documented as v11 invariant in html-hermes-style. Any new style folder must inherit this rule. Sentinel: if the WeChat poster has white bleed past the first viewport, this is the cause.
2. Wrangler --branch=main is required for production apex. Without it, deploy lands in preview-only and the apex <project>.pages.dev continues serving the previous deployment. Always pass --branch=main for prod.
3. CF env-var precedence trap. If CLOUDFLARE_API_TOKEN is set in shell with insufficient scope (e.g. only Worker:Edit, not Pages:Project:Read), wrangler returns 400 on /memberships. Fix: env -u CLOUDFLARE_API_TOKEN -u CF_API_TOKEN -u CLOUDFLARE_ACCOUNT_ID wrangler ... falls back to OAuth in ~/.wrangler/config/default.toml. Encoded into deploy-cf.sh.
4. Empty curl mirror trap. curl -sf URL -o file saved 0-byte file when source returned 200 but body empty (CF revalidation race). Defensive: pair with curl -sL and post-write size check.
5. WeChat 3-layer image-mode rule — none individually sufficient:
   • Format: JPEG > PNG (PNG over ~3 MB or extreme ratio degrades to file-mode)
   • Dimensions: width 1080 = native preview; >2048 forces server recompression; aspect ratio ≤1:8 keeps preview-mode
   • Send method: macOS cmd+C (Preview) → cmd+V (WeChat) forces image-clipboard pipe; drag-drop on macOS WeChat is unreliable for >1 MB. For 企业微信 Bot API, use media/upload?type=image + msgtype=image.
6. Sub-path bundling is BANNED. Each episode = own CF Pages project. the program owner rejected <series-host>/<episode-slug>/ after seeing episode #02 land at luofuli-interview-2026.pages.dev/hermes-agent/ (URL slug ≠ content). Project name pattern: deeplearn-<NN>-<title-slug>.pages.dev.
7. GDrive file naming is <date>_深度学习_第<NN>期_<title> — not <title>_<date>. Date prefix sorts alphabetically by chronology; episode number is part of the structured key. Apply retroactively when discovered.
8. Style-per-episode is load-bearing. Reusing a style across episodes erodes the editorial signature. New episode = new style (stock pick OR custom commission). Past episode style references stay in references/episodes.json for cross-reference, NOT for reuse.
9. scripts/hermes-weixin-send-image.py default path = image-mode. The --as-file flag preserves HD by sending as file (auto-converts to download chip on receiver). Without the flag, the script uses native_image_attempts ladder which auto-picks ≤512 KB preview candidates → image-mode delivery.
10. CF edge cache holds stale entries past max-age=0. After redeploying to remove a sub-path, the apex may serve stale content for ~minutes. Cache-bust query (?bust=$(date +%s)) confirms backend is correct; user-facing apex clears naturally.

Files

• SKILL.md — this spec (single source of truth)
• scripts/render-poster.mjs — Playwright triple-format renderer (4K + mobile + WeChat)
• scripts/deploy-cf.sh — CF Pages dedicated-project deploy with OAuth fallback
• scripts/archive-gdrive.sh — rclone wrapper with structured naming
• scripts/episode-next.sh — query GDrive to discover next episode number
• references/episodes.json — registry of past episodes (#01, #02, ...)
• references/style-decision.md — when to pick stock vs commission custom

Composition with other skills

• deep-learning-skill (existing) — body content generation (3-layer structure, stats, pull-quotes, dropcaps). This pipeline calls it for stage B's content; overrides its template with stage A's style.
• html-style-router — stock style selection for stage A (12 tier styles).
• html-hermes-style — example bespoke style; commission template for new bespoke styles in stage A.
• auto-visual-swarm-review — optional ≥3-round swarm review of HTML poster after stage B (per the program owner rule feedback_auto_swarm_review.md).
• aesthetic-quality-probe (PostToolUse hook) — auto-verifies brand-token adherence after Edit/Write of outputs/reports/**/*.html.

Reference: episode registry

See references/episodes.json for canonical record of #01 and #02. Schema:

{
  "<NN>": {
    "date": "YYYY-MM-DD",
    "title_zh": "...",
    "title_en": "...",
    "style_id": "html-<style>-style",
    "style_path": "skills/shared/html-<style>-style/SPEC.md",
    "cf_url": "https://<project>.pages.dev/",
    "gdrive_prefix": "<date>_深度学习_第<NN>期_<title>",
    "html_path": "outputs/reports/<topic>/<date>-<topic>.html",
    "screenshots": { "4k": "...", "mobile": "...", "wechat": "..." }
  }
}

---

Agent Foundry Team