Gen Social

Objective

Transform the daily AI news report into platform-optimized social media content. Generates ready-to-post copy and adapted images for multiple platforms, with support for multiple accounts and styles per platform.

Phase 1 scope: Generate copy files + images. No auto-publish (future phase).

Supported Platforms

X / Twitter

Xiaohongshu (Little Red Book)

Style Definitions

Custom Personas & Styles

Templates define the voice, tone, and format for each channel. This repo ships with one example template per platform as a starting point:

• X: templates/x/insider.md (Tech Insider persona)
• Xiaohongshu: templates/xiaohongshu/educational.md

To create your own persona or style:

# X persona
cp skills/gen-social/templates/x/insider.md skills/gen-social/templates/x/my-persona.md

# Xiaohongshu style
cp skills/gen-social/templates/xiaohongshu/educational.md skills/gen-social/templates/xiaohongshu/my-style.md

Edit the copy, then set your channel config style field to match the filename (e.g. "style": "my-persona"). Custom templates in these directories are gitignored — your personas stay local.

X Styles

Each X style has a distinct persona — a consistent voice and perspective that makes the account feel like a real person, not a news feed.

Available styles:

Xiaohongshu Styles

Available styles:

Channel Configuration

A channel is the core abstraction: one channel = one platform + one style + one language + content selection rules. Each channel produces an independent set of output files.

Channel Config File

Located at ~/.config/morning-ai/social_channels.json (or path set via SOCIAL_CHANNELS_FILE).

[
  {
    "id": "x_insider_en",
    "platform": "x",
    "style": "insider",
    "lang": "en",
    "items": 3,
    "min_score": 7,
    "image": true,
    "image_aspect": "16:9"
  },
  {
    "id": "x_thread_en",
    "platform": "x",
    "style": "thread",
    "lang": "en",
    "items": 5,
    "min_score": 6,
    "image": true,
    "image_aspect": "16:9"
  },
  {
    "id": "xhs_kepu_zh",
    "platform": "xiaohongshu",
    "style": "educational",
    "lang": "zh",
    "items": 5,
    "min_score": 6,
    "image": true,
    "image_aspect": "3:4",
    "image_count": 4
  },
  {
    "id": "xhs_zhongcao_zh",
    "platform": "xiaohongshu",
    "style": "recommendation",
    "lang": "zh",
    "items": 3,
    "min_score": 7,
    "image": true,
    "image_aspect": "3:4",
    "image_count": 3
  }
]

Channel Fields

Quick Setup (Single Channel via Env Vars)

When only one channel is needed, use env vars instead of a JSON file:

SOCIAL_PLATFORM=x
SOCIAL_STYLE=insider
SOCIAL_LANG=en

This creates a single channel with id {platform}_{style}_{lang}.

Content Selection Rules

For each channel, select items from the daily report data:

1. Filter by score: only items with importance >= min_score
2. Filter by type: if include_types is set, only those content types
3. Filter by KOL voice flag: if kol_voice_only: true is set, only items where is_kol_voice: true
4. Sort by score: highest first
5. Limit: take top N items (from channel items setting or template default)
6. Translate: if source data language differs from channel lang, translate content. Entity names (proper nouns) stay unchanged.

Conditional Channels (Pre-Selection Gate)

This check runs BEFORE the 5 steps above. Some channels only make sense on certain days (e.g. a viral-only channel that should only post when there's a genuinely big story). These channels declare a conditional field — the value tells the pipeline what gate to apply:

When a conditional channel is gated off:

1. Do NOT generate a copy file or images for this channel
2. Still emit a manifest entry, but with a skipped: true field and a skip_reason explaining why (e.g. "max score 7.8 < required 8.5")
3. Required manifest entry shape for skipped conditional channels:

   {
     "id": "xhs_zhongcao_viral_zh",
     "platform": "xiaohongshu",
     "style": "single-topic",
     "skipped": true,
     "skip_reason": "max score 7.8 < required 8.5"
   }
   

4. Downstream publish adapters (e.g. foreman daily-publish-xhs-2-viral) detect skipped: true and mark their run as SKIPPED rather than FAILED.

Why this matters: it lets us schedule a 12:00 viral-only post on the cron without generating empty/low-quality posts on slow news days. The viral channel only fires content when there's something worth firing about.

Copy Generation

For each channel:

1. Read the template: {SKILL_DIR}/skills/gen-social/templates/{platform}/{style}.md
2. Select items using Content Selection Rules above
3. Generate copy following the template's format rules, character limits, and tone
4. Validate character counts against platform limits

Post-Generation Checklist

• Character count within platform limits (280/tweet for X, 20 title + 1000 body for Xiaohongshu)
• Language matches channel lang setting
• Entity names preserved as proper nouns (not translated)
• Source links included where template requires them
• Hashtags/tags present and relevant
• No score numbers displayed in copy (scores are internal only)

Image Generation

Social images reuse the existing image generation infrastructure (lib/image_gen.py) and the same providers (Gemini/GPT/MiniMax).

Image Language Rules

Image text language must match the target platform, not the report source language:

All card titles, bullet points, section headers, and any text rendered on the image must be in the platform's language. Entity names (proper nouns like OpenAI, DeepSeek) remain unchanged.

Persona–Image Style Mapping

Each persona/style has a recommended image style that matches its voice. This overrides the global IMAGE_STYLE setting for social images.

X Styles

Xiaohongshu Styles

IMPORTANT: If the channel config has an explicit image_style field, that takes absolute priority over the defaults above. Do NOT include style labels or content-type annotations in image prompts.

Briefing Per-Category Images

The briefing channel generates additional per-category images beyond the cover, mirroring how gen-infographic handles per-type infographics:

Per-category images:

• Same classic style as cover, 16:9 landscape
• Header: "MorningAI — {Type} Updates" (no date — consistent with infographic per-type images)
• Include ALL qualifying items of that type (not limited to the channel items cap)
• Each card: entity + event + 2-3 bullet points

The recommended style is a default — channels can override via image_style field in the channel config. If not set, the persona's recommended style is used. If no persona mapping exists, falls back to the global IMAGE_STYLE setting.

Platform-Specific Image Adaptation

X images:

• Aspect ratio: 16:9 (landscape) or 1:1 (square)
• All text in English — titles, bullets, headers, everything
• Header: "MorningAI"
• Apply the persona's recommended image style (see mapping above)
• Single image per tweet (or up to 4 for threads)

Xiaohongshu images:

• Aspect ratio: 3:4 (portrait) or 1:1 (square)
• All text in Chinese — titles, bullets, headers, everything
• Header: "MorningAI"
• Apply the style's recommended image style (see mapping above), plus Xiaohongshu-specific adaptations:
  • Larger font sizes for mobile readability
  • Rounded card corners (16px)
  • Emoji-style bullet markers
• Carousel strategy (when image_count > 1):
  • Image 1: Cover overview — top headlines, eye-catching title
  • Image 2-N: Detail pages — 1-2 items per image with expanded bullet points
  • Last image: Follow/subscribe CTA (optional)

Image Prompt Template

{ASPECT} infographic, {HEADER_TEXT} {YYYY-MM-DD}, ALL text content in {LANG}.

Total news items: {N}

News cards (display EXACTLY {N} cards):

Card 1: {Entity name} {Event subject} {Core event verb phrase}
- {Point 1}
- {Point 2}
- {Point 3}

(... list according to actual item count ...)

CRITICAL RULES:
- ALL text on this image MUST be in {LANG} — titles, bullet points, headers, labels, everything
- Entity names are proper nouns (OpenAI, DeepSeek, Cursor) — keep as-is, do NOT translate
- Do NOT display ANY platform names on the image (no "小红书", "Xiaohongshu", "特刊", "Twitter", "X" etc.) — the image is content, not an ad for the platform
- Header text: "{HEADER_TEXT}"
- Each card title MUST include: Entity name + Event subject + Event description
- Display complete titles, do NOT truncate
- Do NOT display score numbers, score badges, or importance markers
- Do NOT invent items not listed
- Display ALL bullet points for each card
- Maximize content area — card titles and bullet points are the primary focus

ANTI-AI-SLOP RULES (mandatory — ref: huashu-design):
- NO purple/lavender/violet gradients — this is the #1 "AI-generated" tell
- NO emoji as bullet markers or icons — use typographic bullets (·, —, or numbered)
- NO rounded-corner cards with colored left border accent — the 2020-2024 AI slop archetype
- NO Inter/Roboto/Arial/system fonts as display — use serif display fonts (Noto Serif, Source Serif, Georgia)
- NO neon glow, glassmorphism, or frosted glass effects
- NO decorative icons per heading — if an icon doesn't carry information, remove it
- NO flat digital color blocks — use paper/material texture feel
- Color restraint: black + cream/white + ONE accent color max. Do NOT invent extra colors
- Typography IS the visual language — let font weight, size, and spacing create hierarchy, not color/shape decoration
- One detail at 120%, everything else at 80% — not uniformly polished, but sharp where it counts
- Describe mood, not pixel coordinates — "feels like a morning broadsheet" beats "title at top, 36pt, centered"

{STYLE_BLOCK}
{PLATFORM_STYLE_ADDON}

Variable substitution:

• {HEADER_TEXT} → "MorningAI" for X, "MorningAI" for Xiaohongshu
• {LANG} → "English" for X, "Chinese" for Xiaohongshu
• {STYLE_BLOCK} → from the persona's recommended image style (see skills/gen-infographic/SKILL.md Style Presets)

{PLATFORM_STYLE_ADDON} for Xiaohongshu — use the matching style variant based on the channel's image_style field:

When image_style = glassmorphism:

Additional Xiaohongshu adaptation (glassmorphism):
- ALL text must be in Chinese (except entity proper nouns)
- Soft gradient background (lavender to pale rose)
- Semi-transparent frosted white cards with backdrop blur
- Use rounded corners (16px) on all cards
- Larger title font (22pt bold) for mobile readability
- Warm accent colors: coral (#FF6B6B), soft pink (#FFB4B4), lavender (#E8EAF6)
- Emoji bullet markers (colored dots or sparkle symbols)
- Clean, fresh, lifestyle-magazine aesthetic
- Generous padding and line spacing for mobile screens

When image_style = newspaper:

Additional Xiaohongshu adaptation (newspaper):
- ALL text must be in Chinese (except entity proper nouns)
- Mood: feels like unfolding a quality morning broadsheet — authoritative, calm, information-dense, zero decoration
- Warm cream background (#FFF8E7) with subtle paper grain texture — NOT flat digital white
- Typography IS the design: bold serif headers (Noto Serif CJK / Source Han Serif / Georgia feel), clean sans-serif body
- Typographic hierarchy through font weight (700 headers, 400 body) and size contrast — NOT through color or shape
- Thin black hairline rules (1px) between sections, like column dividers in a broadsheet
- NO card backgrounds, NO rounded corners, NO shadows, NO colored borders — pure typographic layout
- Color: cream + deep black + crimson (#DC143C) accent ONLY — crimson used sparingly (dateline, one key number, or a thin rule)
- Bullet markers: typographic only (· or —), NEVER emoji
- One signature detail at 120%: the headline typography should feel crafted, like a newspaper masthead
- Generous line spacing (1.6+) for mobile readability, but information-dense — fill the space with content, not decoration
- DO NOT use gradients, frosted glass, glow effects, or any "digital" aesthetic
- The image should look like it was typeset, not generated

When image_style = cover-hook: (FIRST image of XHS carousel ONLY — this is the discovery-feed thumbnail)

Additional Xiaohongshu adaptation (cover-hook):
- ALL text must be in Chinese (except entity proper nouns)
- Mood: a newspaper EXTRA edition headline — urgent, bold, one statement that demands attention, but with editorial class, not digital loudness
- 3:4 single card, super-large title text occupying ≥ 50% of canvas
- Title text = the post title or its strongest 1-line variant (e.g., "Cursor 估值飙到 500 亿")
- ONE single subtitle line below with the most striking concrete data point (e.g., "3 年达 20 亿 ARR · 史上最快 B2B")
- Newspaper headline poster aesthetic:
  - Warm cream (#FFF8E7) with bold black serif title, OR deep charcoal (#1A1A1A) with cream/off-white serif title
  - Bold serif typography (Noto Serif CJK / Source Han Serif / Georgia feel), weight ≥ 700
  - Paper grain texture — NOT flat digital color blocks
- Thin crimson (#DC143C) hairline rule above or below the headline — the ONE signature detail at 120%
- Optional: small "MorningAI" masthead at very top in small caps, like a newspaper nameplate — restrained, not decorative
- NO bullet points, NO emoji, NO icons, NO small annotations — single statement only
- NO body text, NO list of items, NO date stamps — only the headline + one subtitle line
- NO gradients, NO glow, NO frosted glass, NO rounded corners — this is typeset, not generated
- The cover should look like a broadsheet front page that was photographed, not a Canva template

When image_style = classic:

Additional Xiaohongshu adaptation (classic):
- ALL text must be in Chinese (except entity proper nouns)
- Off-white background (#F5F5F0), clean editorial magazine layout
- Card-based layout with subtle shadows
- Navy (#1B2A4A), coral (#E8634A), teal (#2A9D8F) accent palette
- Use rounded corners (8px) on cards
- Larger title font for mobile readability
- Structured, authoritative visual hierarchy
- Generous padding and line spacing for mobile screens

{PLATFORM_STYLE_ADDON} for X (append to any base style):

Additional X/Twitter adaptation:
- ALL text must be in English (except entity proper nouns)
- Optimized for timeline scroll — key info visible at small preview size
- High contrast text for readability on mobile

Generating Images

Use the same methods as Step 4 (gen-infographic):

Option A — Native tool (if supported): Generate each image using built-in image generation capability.

Option B — Python script batch mode: Build a manifest JSON and run:

cd {SKILL_DIR} && python3 skills/gen-infographic/scripts/gen_infographic.py --batch {CWD}/social/manifest_images.json

Manifest entries support "aspect_ratio": "3:4" for Xiaohongshu images.

Cover-Hook for Xiaohongshu Carousels

For Xiaohongshu channels with image_count >= 2 AND style recommendation or single-topic, the first image of the carousel is the discovery-feed thumbnail and MUST use the cover-hook style addon (regardless of the channel's regular image_style):

• Image 1 → use {PLATFORM_STYLE_ADDON} for cover-hook. Content = post title + ONE strongest data point. No bullet points, no list of items.
• Image 2..N → use the channel's regular image_style (e.g., newspaper). Content = the per-item news cards as before.

This applies because the first image is what stops the scroll in the discovery feed; an information-dense newspaper card is unreadable as a thumbnail. Channels that already declare image_style: cover-hook explicitly should apply it to ALL images (rare — typically reserved for single-image posts).

Output Files

All output goes to {CWD}/social/ directory.

Naming Convention

social/
├── social_{DATE}_{channel_id}.md              # Copy file
├── social_{DATE}_{channel_id}_cover.png       # Single/cover image
├── social_{DATE}_{channel_id}_{N}.png         # Carousel image N
└── social_{DATE}_manifest.json                # Output index

Manifest Format

MANDATORY SCHEMA — DO NOT IMPROVISE FIELD NAMES OR PATHS

The manifest is consumed by downstream tools that match on exact field names and assume flat filenames. Past LLM-generated output has drifted in ways that silently break consumers (channel_id instead of id, copy instead of copy_file, paths prefixed with social/). This section is the contract — produce it byte-for-byte as specified.

Required top-level shape:

• channels MUST be a JSON array (list). Never a dict keyed by id.

Required per-channel fields — use these exact names, no aliases:

Path rules — applies to copy_file and every entry in images:

• Write the bare filename as it appears under social/ (e.g. social_2026-04-18_xhs_zixun_zh.md).
• Do NOT include any directory component — no social/, no ./, no absolute path.
• Wrong: "copy_file": "social/social_2026-04-18_xhs_zixun_zh.md" ❌
• Right: "copy_file": "social_2026-04-18_xhs_zixun_zh.md" ✅

Canonical example (copy this shape exactly):

{
  "date": "2026-04-14",
  "channels": [
    {
      "id": "x_insider_en",
      "platform": "x",
      "style": "insider",
      "lang": "en",
      "copy_file": "social_2026-04-14_x_insider_en.md",
      "images": ["social_2026-04-14_x_insider_en_cover.png"],
      "items_used": 3
    },
    {
      "id": "xhs_kepu_zh",
      "platform": "xiaohongshu",
      "style": "educational",
      "lang": "zh",
      "copy_file": "social_2026-04-14_xhs_kepu_zh.md",
      "images": [
        "social_2026-04-14_xhs_kepu_zh_1.png",
        "social_2026-04-14_xhs_kepu_zh_2.png",
        "social_2026-04-14_xhs_kepu_zh_3.png",
        "social_2026-04-14_xhs_kepu_zh_4.png"
      ],
      "items_used": 5
    },
    {
      "id": "xhs_zhongcao_viral_zh",
      "platform": "xiaohongshu",
      "style": "single-topic",
      "lang": "zh",
      "skipped": true,
      "skip_reason": "max score 7.8 < required 8.5"
    }
  ]
}

Forbidden variants (all of these have been observed in real output and break downstream consumers — do not produce them):

// ❌ WRONG — "channel_id" is not a recognized field
{ "channel_id": "xhs_zixun_zh", ... }

// ❌ WRONG — "copy" is not a recognized field
{ "id": "xhs_zixun_zh", "copy": "...", ... }

// ❌ WRONG — paths must be bare filenames
{ "copy_file": "social/social_2026-04-18_xhs_zixun_zh.md",
  "images": ["social/social_2026-04-18_xhs_zixun_zh_1.png"] }

// ❌ WRONG — channels must be a list, not a dict keyed by id
{ "channels": { "xhs_zixun_zh": { ... } } }

Before writing the manifest file, mentally diff your output against the canonical example above. If any field name, structure, or path shape differs, fix it before writing.

Workflow Summary

1. Check if social content is enabled (SOCIAL_ENABLED=true)
2. Load channel configuration (JSON file or env vars)
3. For each channel: a. Conditional gate: if the channel has a conditional field, evaluate the gate. If it fails, write a manifest entry with skipped: true + skip_reason and continue to the next channel — do NOT generate copy or images. b. Read the channel's template file c. Select top items from report data (score filter + item limit) d. Generate copy following template rules and character limits e. Write copy to social/{DATE}_{channel_id}.md f. If image: true — build image prompts and generate platform-adapted images. For XHS carousels with style recommendation or single-topic, the FIRST image uses the cover-hook style addon; subsequent images use the channel's regular image_style. g. Write images to social/{DATE}_{channel_id}_{N}.png
4. Write manifest to social/{DATE}_manifest.json

Notes

• Channels are fully independent — each reads from the same daily data but produces separate output
• Multiple channels can target the same platform with different styles, languages, or content focus
• Image language follows the platform: X images are always English, Xiaohongshu images are always Chinese — regardless of the channel lang setting for copy
• Image style follows the persona: each persona/style has a recommended image style that matches its voice (see Persona–Image Style Mapping). Channels can override via image_style field
• All content must respect platform character limits — validation is mandatory
• Entity names are proper nouns and must NOT be translated regardless of lang setting
• When the same item appears in multiple channels, each channel generates its own adapted version independently