Operating Conventions
- CLI: deterministic operations go through the
wewritecommand (must be in PATH; if missing, guide touv tool install wewrite, or in the repositorybash install.sh）。 - {home}: user state directory =
$WEWRITE_HOMEor~/.wewrite（wewrite homecan be checked). config/style/history/playbook/output/exemplars are all in {home}, not in the repository; state paths in references documents follow the same convention. 读取: <路径>= actually finish reading the file with a file-reading tool before continuing, not a comment.- references/: this skill comes with
{skill_dir}/references/;{skill_dir}inside references documents refers to this skill's directory. - Task state: read and update via
wewrite run show/update/step; do not directly modify the state file.
Prerequisites
{home}/style.yamlexists → extracttopics、content_style. Does not exist → ask the user "what main directions does your WeChat official account write about?" Use the answer as a temporary substitute for topics, and prompt that wewrite-style can be used to complete the full setup.
2.1 Trending Topic Scraping
wewrite hotspots --limit 30
Fallback: script errors → WebSearch "today's trending topics {first vertical category in topics}"
2.1b High-Frequency Demand Reference
For topics 's first 1-2 vertical category words, search once each for WeChat official account articles from the past week:
wewrite search-articles "{垂类词}" -n 15 -t 2
Produce a list of WeChat official account articles in the same vertical category from the past week (title/summary/publish time/account), as a "verified content demand" signal for use in 2.3. Sogou results do not include read counts, so they can only indicate the frequency of the same topic appearing, and cannot claim an article is a "viral hit".
Fallback: command errors / empty results (Sogou anti-scraping rate limiting) → skip this step, do not block, and the topic composition falls back to default.
2.2 Historical Analysis + SEO
读取: {home}/history.yaml（不存在则跳过）
wewrite seo --json {关键词}
Historical analysis (when stats data is available):
- Count which
framework's articles performed best (read count/share rate) → weight when recommending frameworks - Count which
enhance_strategy's articles performed best → reference when selecting enhancement strategies - Keywords already written in the past 7 days get a score reduction (deduplication)
Fallback: SEO script errors → LLM judgment; history has no stats → skip performance analysis, only do deduplication
2.3 Generate Topic Ideas
读取: {skill_dir}/references/topic-selection.md
Generate 10 topic ideas, of which:
- 5-6 trending topic ideas: based on the trending topics from 2.1, scored according to topic-selection.md rules
- 2-3 high-frequency angle topic ideas(only when 2.1b has output): based on articles on the same topic from the past week, scored according to topic-selection.md rules, marked "high-frequency demand"; when 2.1b has no output, the share goes back to trending topics
- 2-3 evergreen topic ideas: not dependent on trending topics, generated from the user's
topicsdomain as long-tail content (tutorials/methodologies/experience summaries/tool recommendations), marked as "evergreen". Suitable for users whose content_style is practical/knowledge-based or review-based
Each topic idea includes title, score, click-through potential, SEO friendliness, and recommended framework.
- Auto mode → select the highest score
- Interactive mode / standalone activation → show all, wait for the user to choose
Done
Use wewrite run update --patch to write topic.title、topic.keywords、topic.source
("hot topic capture" / "high-frequency demand" / "evergreen"),topic.framework_hint, then
wewrite run step topic completed.
When activated standalone and the user only wants a list of topics, just show 10 topics; after the user chooses, write to the state and prompt
"you can directly say 'just write this' to enter writing (wewrite-write)".