Creator

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)

Language Adaptation: All UI text follows the user's input language. Chinese input → Chinese output. English input → English output. Mixed → follow dominant language.

Use AskUserQuestion for every multiple-choice step. One question at a time. Wait for the answer. After template is selected and input is understood, show a confirmation summary and wait for explicit approval before executing the pipeline.

API Key Check at Confirmation Gate: If the pipeline includes any remote API call (image-gen, content-parser, tts), check authentication before proceeding. For CLI-based calls (image-gen, TTS), run listenhub auth login if not authenticated. For content-parser calls, configure LISTENHUB_API_KEY (see content-parser/SKILL.md § Authentication). Pure text-only pipelines (e.g., topic → narration script without TTS) can proceed without authentication.

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" 50 chars, not a URL/path | Use directly as material |
| Topic/keywords | Short text (/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 "
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 "" \
  --aspect-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

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: joeseesun
• Source: joeseesun/qiaomu-cut-skill
• License: MIT

Install and usage instructions live in the source repository linked above.