When to Use

• User wants a full content package for a specific platform (WeChat article, Xiaohongshu post, narration script)
• User says "帮我写篇公众号", "小红书图文", "口播稿", "create content"
• User provides a URL/text/topic and wants it turned into platform-ready content with images

When NOT to Use

• User wants a single image without a content workflow → use image-gen directly
• User wants a single TTS audio → use tts directly
• User wants to transcribe audio → use asr directly
• User wants a podcast episode → use podcast directly
• User wants to extract content from a URL without further processing → use content-parser directly

Creator is for multi-step content production that combines writing + media generation into a platform-ready package.

Purpose

Generate platform-specific content packages by orchestrating existing skills. Input: topic, URL, text, or audio/video file. Output: a folder with article/script, images, and metadata — ready to publish.

Hard Constraints

• Use listenhub CLI commands for image-gen and TTS. Use curl for content-parser (see content-parser/SKILL.md § API Reference).
• Always read config following shared/config-pattern.md before any interaction
• Follow shared/cli-patterns.md for polling, errors, and interaction patterns
• Never save files to ~/Downloads/ or .listenhub/ — save content packages to the current working directory
• JSON parsing: use jq only (no python3, awk)

Step -1: API Key Check

Deferred. API key is checked at the confirmation gate (Step 4) only when the pipeline requires remote API calls. See Hard Constraints above.

Step 0: Config Setup

Follow shared/config-pattern.md Step 0 (Zero-Question Boot).

If file doesn't exist — silently create with defaults and proceed:

mkdir -p ".listenhub/creator" ".listenhub/creator/styles"
cat > ".listenhub/creator/config.json" << 'EOF'
{"outputMode":"download","language":null,"preferences":{"wechat":{"history":[]},"xiaohongshu":{"mode":"both","history":[]},"narration":{"defaultSpeaker":null,"history":[]}}}
EOF
CONFIG_PATH=".listenhub/creator/config.json"
CONFIG=$(cat "$CONFIG_PATH")

User style preferences are stored as markdown files in .listenhub/creator/styles/:

• .listenhub/creator/styles/wechat.md
• .listenhub/creator/styles/xiaohongshu.md
• .listenhub/creator/styles/narration.md

These files are plain markdown — one directive per line. If the file does not exist, no custom style is applied. Users can edit these files directly.

Note: outputMode defaults to "download" (not the usual "inline") because creator always produces multi-file output folders that must be saved to disk.

If file exists — read config silently and proceed:

CONFIG_PATH=".listenhub/creator/config.json"
[ ! -f "$CONFIG_PATH" ] && CONFIG_PATH="$HOME/.listenhub/creator/config.json"
CONFIG=$(cat "$CONFIG_PATH")

Setup Flow (user-initiated reconfigure only)

Only when user explicitly asks to reconfigure. Display current settings:

当前配置 (creator)：
  输出方式：{outputMode}
  小红书模式：{both / cards / long-text}

Ask:

1. outputMode: Follow shared/output-mode.md § Setup Flow Question.
2. xiaohongshu.mode: "小红书默认模式？"
   • "图文 + 长文（both）"
   • "仅图文卡片（cards）"
   • "仅长文（long-text）"

Interaction Flow

Step 1: Understand Input

The user provides input along with their request. Classify the input:

Style reference detection: If the user's prompt contains keywords like "参考", "风格", "照着…写", "style", "reference", the associated input (file path / URL / pasted text) should be classified as a style reference rather than content material. A single request may contain both material and a style reference — classify them separately. If only a style reference is provided with no material or topic, this is a standalone style learning request (see Step 2.5).

For URL (audio/video) inputs:

1. Download to /tmp/creator-{slug}.{ext} using curl -L -o
2. Check coli is available: which coli 2>/dev/null && echo yes || echo no
3. If coli missing: inform user to install (npm install -g @marswave/coli), ask them to paste text instead
4. Transcribe: coli asr -j --model sensevoice "/tmp/creator-{slug}.{ext}"
5. Extract text from JSON result
6. Cleanup: rm "/tmp/creator-{slug}.{ext}"

For URL (web/article) inputs: Content-parser will be called during pipeline execution (after confirmation).

Step 2: Template Matching

If the user specified a platform in their prompt, match directly:

• "公众号", "wechat", "微信" → wechat
• "小红书", "xiaohongshu", "xhs" → xiaohongshu
• "口播", "narration", "脚本" → narration

If no platform was specified, ask via AskUserQuestion:

Question: "Which content template?" / "用哪个创作模板？" Options (adapt language to user's input):

• "WeChat article (公众号长文)" — Long-form article with AI illustrations
• "Xiaohongshu (小红书)" — Image cards + long text post
• "Narration script (口播稿)" — Spoken script with optional audio

Step 2.5: Topic Assistance

This step runs only when the user's input is a topic or keywords (short text <50 chars, no URL/path). Skip if user provided a URL, file, or substantial text.

1. Read the selected platform's methodology.md:

   • WeChat: creator/templates/wechat/methodology.md
   • Xiaohongshu: creator/templates/xiaohongshu/methodology.md
   • Narration: creator/templates/narration/methodology.md

2. Evaluate the topic using the three-circle Venn model:

   • 用户的专业领域 (creator's expertise)
   • 读者的普遍兴趣 (reader interest)
   • 当下的时间节点 (current timing/relevance)

3. Run HKR quality filter:

   • H (Happy): 足够有趣、有悬念？
   • K (Knowledge): 有信息量？看完能学到新东西？
   • R (Resonance): 能戳中情绪？让人"对对对我也这么想"？

4. If topic scores ≥2 of 3 HKR criteria: proceed with the topic.

5. If topic scores <2: proactively suggest 2-3 alternative angles to the user via AskUserQuestion.

6. If topic is vague: ask for more specifics — key points, personal experiences, what excites or frustrates them.

Step 3: Style Extraction (if style reference provided)

This step runs only when the user provided a style reference in Step 1. If no style reference was detected, skip to Step 3b.

Read the reference content:

• Local file → Read tool
• URL → content-parser API (requires API key)
• Pasted text → use directly

Analyze and extract style directives:

AI reads the reference content and extracts 3-5 concrete style directives. Focus on observable patterns:

• Sentence length and paragraph structure
• Tone and register (formal/casual, first/third person)
• Use of rhetorical devices (questions, lists, bold, quotes)
• Vocabulary level and domain jargon
• Formatting habits (heading style, emoji usage, whitespace)

Present to user for confirmation:

从参考文章中提炼了以下风格特征：

  1. {directive 1}
  2. {directive 2}
  3. {directive 3}
  ...

你可以修改或删除其中的条目。确认后本次生成会应用这些规则。

Wait for user confirmation. The confirmed directives become sessionStyle — applied to this generation only.

After user confirms the style directives, proactively ask whether to persist:

要将这些风格规则保存吗？（保存后每次生成{platform}内容都会应用）

If yes → write to .listenhub/creator/styles/{platform}.md. If no → only apply to this generation.

Standalone style learning: If the user only provided a style reference without material/topic (e.g., "学习一下这篇文章的风格"), run the extraction above, then persist directly to .listenhub/creator/styles/{platform}.md without asking — the user's intent to save is already explicit. Confirm with a brief message: "已保存到 styles/{platform}.md". Do not proceed to content generation.

Step 3a: Prototype Classification

Read the selected platform's prototype file:

• WeChat: creator/templates/wechat/article-prototypes.md
• Xiaohongshu: creator/templates/xiaohongshu/content-prototypes.md
• Narration: creator/templates/narration/script-prototypes.md

Based on the user's material/topic, auto-match the best-fit prototype using the matching heuristics table in the prototype file.

Present the recommendation to the user via AskUserQuestion:

Question: "这篇内容最适合哪种写法？" / "Which content prototype fits best?" Options: [list all prototypes for the platform, recommended one first with "(Recommended)" suffix]

The selected prototype determines the narrative structure and L3-5 review criteria for writing.

Step 3b: Preset Selection (if applicable)

If the selected template uses illustration or card presets and the mode requires images, the preset MUST be chosen before the confirmation gate so it can be displayed in the summary.

Skip this step entirely for:

• Narration template (no visual presets)
• Xiaohongshu with preferences.xiaohongshu.mode = "long-text" (no cards or images generated)

Otherwise:

1. Read the template's preset section to get available presets and the topic-matching table.
2. If the user already specified a preset in their prompt (e.g., "用水彩风格"): use that preset directly.
3. If not specified: ask the user via AskUserQuestion. Output a one-line hint first: "配图风格可以随时换，先选一个开始吧". List all available presets with their Chinese labels (from frontmatter label field). Use the topic-matching table to put the most relevant option first (marked "Recommended"), but always let the user choose.

Step 4: Confirmation Gate

Check API key if the pipeline needs remote APIs:

• WeChat template always needs image-gen → requires API key
• Xiaohongshu cards mode needs image-gen → requires API key
• Xiaohongshu long-text only → no API key needed
• Narration without TTS → no API key needed
• Web/article URL input → needs content-parser → requires API key (audio/video URLs use local coli asr, no API key needed)

If API key required and missing: for CLI-based calls, run listenhub auth login. For content-parser calls, configure LISTENHUB_API_KEY (see content-parser/SKILL.md § Authentication).

Show confirmation summary:

准备生成内容：

  模板：{WeChat article / Xiaohongshu / Narration}
  输入：{topic description / URL / text excerpt...}
  输出目录：{slug}-{platform}/
  需要 API 调用：{content-parser, image-gen, ...}
  风格偏好：{styles/{platform}.md 已配置 / 使用默认风格}
  配图/卡片预设：{preset label / 不适用}
  文章/内容原型：{selected prototype name}
  本次风格参考：{M条来自参考文章 / 无}

确认开始？

Wait for explicit "yes" / confirmation before proceeding.

Step 5: Execute Pipeline

Read the selected template file and execute:

# The template file path
TEMPLATE="creator/templates/$PLATFORM/template.md"
STYLE="creator/templates/$PLATFORM/style.md"

For URL inputs — extract content first:

# Submit content extraction
RESPONSE=$(curl -sS -X POST "https://api.marswave.ai/openapi/v1/content/extract" \
  -H "Authorization: Bearer $LISTENHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Source: skills" \
  -d "{\"source\":{\"type\":\"url\",\"uri\":\"$INPUT_URL\"}}")
TASK_ID=$(echo "$RESPONSE" | jq -r '.data.taskId')

Then poll in background. Run this as a separate Bash call with run_in_background: true and timeout: 600000 (per shared/cli-patterns.md). The polling loop itself runs up to 300s (60 polls × 5s); timeout: 600000 is set higher at the tool level to give the Bash process headroom beyond the poll budget:

# Run with: run_in_background: true, timeout: 600000
TASK_ID="<id>"
for i in $(seq 1 60); do
  RESULT=$(curl -sS "https://api.marswave.ai/openapi/v1/content/extract/$TASK_ID" \
    -H "Authorization: Bearer $LISTENHUB_API_KEY" \
    -H "X-Source: skills" 2>/dev/null)
  STATUS=$(echo "$RESULT" | tr -d '\000-\037\177' | jq -r '.data.status // "processing"')
  case "$STATUS" in
    completed) echo "$RESULT"; exit 0 ;;
    failed) echo "FAILED: $RESULT" >&2; exit 1 ;;
    *) sleep 5 ;;
  esac
done
echo "TIMEOUT" >&2; exit 2

Extract content: MATERIAL=$(echo "$RESULT" | jq -r '.data.data.content')

If extraction fails: tell user "URL 解析失败，你可以直接粘贴文字内容给我" and stop.

Then follow the platform template — read template.md and execute each step. The template specifies the exact writing instructions and API calls. See creator/templates/{platform}/template.md for template contents.

Writing engine integration: Each platform's template.md now includes writing-engine references and a self-review loop. The template handles loading writing-engine/ files, applying the selected prototype's narrative structure, and running L1-L4 quality review after writing. See each platform's template.md for details.

Style application: When writing content, apply style directives in this priority order (higher overrides lower):

1. sessionStyle — directives from the current style reference (Step 3), if any
2. .listenhub/creator/styles/{platform}.md — persisted user style directives (if file exists)
3. templates/{platform}/style.md — baseline platform style

For image generation (called by wechat and xiaohongshu templates):

RESPONSE=$(listenhub image create \
  --prompt "<generated prompt>" \
  --aspect-ratio "<ratio>" \
  --json)

BASE64_DATA=$(echo "$RESPONSE" | jq -r '.candidates[0].content.parts[0].inlineData.data // .data')
# macOS uses -D, Linux uses -d (detect platform)
if [[ "$(uname)" == "Darwin" ]]; then
  echo "$BASE64_DATA" | base64 -D > "{output-path}/{filename}.jpg"
else
  echo "$BASE64_DATA" | base64 -d > "{output-path}/{filename}.jpg"
fi

On 429: exponential backoff (wait 15s → 30s → 60s), retry up to 3 times. On failure after retries: skip this image, annotate in output summary.

Generate images sequentially (not parallel) to respect rate limits.

For TTS (called by narration template when user wants audio):

listenhub tts create --text "$(cat /tmp/lh-content.txt)" --speaker "$SPEAKER_ID" --json \
  | jq -r '.data' | base64 -D > "{slug}-narration/audio.mp3"

Step 6: Assemble Output

Create the output folder and write all files:

SLUG="{topic-slug}"
OUTPUT_DIR="${SLUG}-{platform}"
# Dedup folder name
i=2; while [ -d "$OUTPUT_DIR" ]; do OUTPUT_DIR="${SLUG}-{platform}-${i}"; i=$((i+1)); done
mkdir -p "$OUTPUT_DIR"

Write content files per template spec. Then write meta.json:

{
  "title": "...",
  "slug": "...",
  "platform": "wechat|xiaohongshu|narration",
  "date": "YYYY-MM-DD",
  "tags": ["...", "..."],
  "summary": "..."
}

Step 7: Present Result

✅ 内容已生成！保存在 {OUTPUT_DIR}/

📄 {main files list}
🖼️ images/ — N 张配图（如有）
📋 meta.json — 标题、标签、摘要

(Adapt language to user's input language per Hard Constraints.)

Step 8: Update Preferences

Record this generation in history:

NEW_CONFIG=$(echo "$CONFIG" | jq \
  --arg platform "$PLATFORM" \
  --arg date "$(date +%Y-%m-%d)" \
  --arg topic "$TOPIC" \
  '.preferences[$platform].history = (.preferences[$platform].history + [{"date": $date, "topic": $topic}])[-5:]')
echo "$NEW_CONFIG" > "$CONFIG_PATH"

Keep only the last 5 history entries per platform.

Note: cardStyle from the spec is deferred — not implemented in V1 config. Can be added later when card style customization is needed.

Manual Style Tuning

Adding style directives:

If the user says "记住：{style directive}" or "remember: {style directive}":

1. Detect which platform it applies to (from context or ask)
2. Append the directive as a new line to .listenhub/creator/styles/{platform}.md (create the file if it doesn't exist)

This also applies after Step 3 (Style Extraction): if the user says "记住这个风格" after reviewing extracted directives, write all confirmed directives to .listenhub/creator/styles/{platform}.md.

Resetting style:

If the user says "重置风格偏好" or "reset style":

1. Ask which platform (or all)
2. Delete .listenhub/creator/styles/{platform}.md

API Reference

• Authentication: shared/cli-authentication.md
• Image generation: CLI: listenhub image create (see shared/cli-patterns.md)
• Content extraction: content-parser/SKILL.md § API Reference (Inlined)
• TTS (text-to-speech): CLI: listenhub tts create (see shared/cli-patterns.md)
• Speaker selection: shared/speaker-selection.md
• Config pattern: shared/config-pattern.md
• Common patterns (polling, errors): shared/cli-patterns.md
• Output mode: shared/output-mode.md

Composability

• Invokes: content-parser (URL extraction), image-gen (illustrations/cards), tts (narration audio), asr (audio/video transcription via coli)
• Invoked by: standalone — user triggers directly
• Templates: creator/templates/{wechat,xiaohongshu,narration}/template.md define per-platform pipelines
• Style guides: creator/templates/{wechat,xiaohongshu,narration}/style.md define per-platform writing tone