Title Tag Cover Generator

Title Tag Cover Generator

Use this skill to turn one article into a publishable Chinese channel material package. Always run it in two stages with human review between them.

Required Flow

1. Stage 1: article -> title candidates + tag strategy + AI recommendations.
2. Stop at Review Gate 1. Do not generate a cover until the user confirms the title.
3. Stage 2: confirmed title -> cover prompt -> image provider or mock -> final package.
4. Stop at Review Gate 2 after cover generation for user inspection.

Never auto-publish, connect CRM, generate video, bypass title review, or put API keys in files or logs.

Responsibility Boundary

This skill owns article analysis, platform titles, platform tags, cover prompt/image generation, and final package assembly. It does not own content publishing, CRM integration, database writes, video production, or final editorial approval.

Inputs

For local test runs, expect:

• input/article.txt
• input/title_reference.md
• input/tag_reference.md
• optional input/cover_references/
• optional cover rules supplied by the user

For production runs, the article may come from another agent. Optional metadata: target platforms, brand or column name, reference cover image, and cover style rules.

Stage 1

Run:

python scripts/run_stage1.py --article input/article.txt --title-reference input/title_reference.md --tag-reference input/tag_reference.md --out outputs/stage1

Then review and improve the generated files manually if needed using the reference documents:

• article_analysis.md
• title_options.md
• selected_title_recommendation.md
• tag_strategy.md
• selected_tags_recommendation.md
• stage1_report.md

Title requirements:

• Generate titles from the article; do not require a user-provided title.
• Use the title reference document as the strategy source.
• Provide candidates for Xiaohongshu, Douyin/Shipinhao, and WeChat article.
• Recommend one title per platform and briefly explain why.
• Prefer high-cognition formulas such as "not A but B", "from A to B", "the stronger A gets, the more important B becomes", and "the real risk is not A but B".

Tag requirements:

• Use the tag reference document as the strategy source.
• Keep total tags at 5-8 per platform.
• Xiaohongshu: search, long-tail terms, concrete scenarios.
• Douyin/Shipinhao: algorithm distribution, precise audience or circle terms.
• Sort by traffic precision, avoid duplicate tags and hollow broad terms.

At the end of Stage 1, ask the user to confirm:

• which title to use,
• which words should be highlighted on the cover,
• whether tags pass,
• whether title revisions are needed.

Stage 2

Only start after the user confirms a title. Stage 2 is intentionally separate.

Inputs:

• confirmed title,
• optional subtitle,
• line breaks,
• green highlight words,
• cover reference image or rules,
• config/image_api_config.example.json copied to a real local config if needed.

Required outputs:

• outputs/cover/confirmed_title.json
• outputs/cover/cover_prompt.md
• outputs/cover/cover_prompt.json
• outputs/cover/cover_image.png
• outputs/cover/cover_generation_report.md
• outputs/final/channel_content_package.json
• outputs/final/production_report.md

Run:

python scripts/run_stage2.py --confirmed-title "<confirmed title>" --line-breaks "line 1|line 2|line 3" --highlight-words "word1,word2"

Cover rules:

• Generate a complete 3:4 cover image or mock placeholder.
• Include the confirmed title text on the cover.
• Highlight key words in green.
• Preserve reference-like spacing, margins, hierarchy, and bottom image area when a reference is provided.
• Leave top space for later tags and bottom space for logo.
• Do not add unrelated text or change the title meaning.

Image API Rules

Use mock mode until a real provider is configured. Mock mode only validates the workflow and should not be treated as final visual quality. Read API keys only from the environment variable named by api_key_env. Never print or store the key.

Final Package

channel_content_package.json must include article summary, core angle, platform title pools, selected titles, tags, cover metadata, publish copy, and remaining human review items.