post-creator

name	post-creator
description	Create reading-hub posts from YouTube videos, web article URLs, pasted text, existing markdown, or user drafts. Use when the user wants to add/archive/save content as a reading-friendly article, says "turn this into a post", "发到 reading-hub", "转成朗读文章", "加一篇文章", or provides a source and asks for a reading post. Writes `src/content/posts/<slug>.md` with Astro-valid frontmatter and source-faithful body content.
allowed-tools	["Read","Write","Bash","Glob","AskUserQuestion"]

Reading Hub Post Creator

Create one Astro content post at src/content/posts/<slug>.md. The post must parse, preserve the source's substance, and keep the source's English voice. For YouTube, the cleaned subtitle transcript is the sole authority for the body.

Steps

1. Classify source. Identify exactly one branch: YouTube, web article URL, raw text/existing markdown, or user draft. Ask only for a missing title or source needed to proceed. Completion criterion: branch, title basis, and source URL/file/text are known.
2. Build clean body. Convert source to markdown using the branch rules below. Completion criterion: body has no extraction artifacts and preserves the source's substantive content and voice.
3. Prepare frontmatter. Use today's date from date +%Y-%m-%d, a valid slug, and only fields supported by src/content.config.ts. Completion criterion: frontmatter is YAML-safe, has title, and tags is a YAML list.
4. Write post. Save only under src/content/posts/<slug>.md; if the file exists, ask before overwriting. Completion criterion: file exists at that path and no unrelated files were changed.
5. Run chunk mining. For every English or bilingual post, run english-chunks-miner after the body is written and before validation. Chinese-only posts omit chunks. Completion criterion: English/bilingual frontmatter includes chunks: or the post is correctly marked Chinese-only.
6. Validate. Run npm run build from the reading-hub repo root. If it fails because of this post, fix and rerun once. Completion criterion: build passes, or the remaining failure is unrelated and reported.
7. Finish. Report the linked path and preview command: cd reading-hub && npm run dev. Do not commit or push unless asked.

Frontmatter

Supported fields: title, date, description, tags, source, cover, speaker, format, language, purpose, chunks.

Use this layout. Omit unknown fields; never invent speaker, source, or quotes.

---
title: "<title>"
date: 2026-04-22
description: "<one-sentence summary>"
source: https://...
cover: https://img.youtube.com/vi/<videoId>/maxresdefault.jpg
speaker: <name, role>
format: <format>
language: <English | Chinese | Bilingual>
purpose: Read-aloud article
tags:
  - tag-a
  - tag-b
---

Rules:

• date is the generation date, not source publish date. Backfilled old content should appear as newly added.
• Quote strings containing : or —.
• Slug is lowercase ASCII, hyphen-separated, no punctuation except -. For Chinese titles, use pinyin or an English topic slug.
• For every English or bilingual post, run english-chunks-miner after the body is written and before validation. Chinese-only posts omit chunks.

Branch Rules

YouTube

Work from the reading-hub repo root. Store intermediates in tmp/youtube-clips/<videoId>/.

1. Fetch subtitles and metadata with yt-dlp:

mkdir -p tmp/youtube-clips/<videoId>
yt-dlp --skip-download --write-auto-subs --sub-lang en --sub-format vtt \
  --cookies-from-browser chrome \
  -o 'tmp/youtube-clips/%(id)s/%(id)s.%(ext)s' <url>
yt-dlp --skip-download --print "%(title)s|%(uploader)s|%(upload_date)s" \
  --cookies-from-browser chrome <url>

2. Convert the VTT to a sentence-level SRT, flatten its text into tmp/youtube-clips/<videoId>/<videoId>.flat.txt, and run rolling-caption deduplication into tmp/youtube-clips/<videoId>/<videoId>.clean.txt whenever repeated display fragments remain:

python3 .github/skills/youtube-clipper/scripts/vtt_to_shadow_srt.py \
  tmp/youtube-clips/<videoId>/<videoId>.en.vtt \
  tmp/youtube-clips/<videoId>/<videoId>.srt
python3 - <<'PY'
from pathlib import Path

srt_path = Path("tmp/youtube-clips/<videoId>/<videoId>.srt")
flat_path = Path("tmp/youtube-clips/<videoId>/<videoId>.flat.txt")

lines = []
for raw in srt_path.read_text(encoding="utf-8").splitlines():
    line = raw.strip()
    if line and not line.isdigit() and "-->" not in line:
        lines.append(line)

flat_path.write_text(" ".join(lines) + "\n", encoding="utf-8")
PY
python3 .github/skills/youtube-clipper/scripts/dedupe_rolling_captions.py \
  tmp/youtube-clips/<videoId>/<videoId>.flat.txt \
  tmp/youtube-clips/<videoId>/<videoId>.clean.txt

If English captions are absent or unusable, stop and report the failure. The flattened subtitle text lives at tmp/youtube-clips/<videoId>/<videoId>.flat.txt; after dedupe, treat tmp/youtube-clips/<videoId>/<videoId>.clean.txt as the canonical cleaned transcript. Dedupe is mandatory whenever sliding-window repeats remain.

3. Keep a cleaned, sentence-level transcript:

• Preserve the speaker's wording, sequence, idioms, complete thoughts, [applause], [laughter], and Q&A turns.
• Only remove display repeats and subtitle artifacts; correct obvious transcription errors; add punctuation, capitalization, and paragraphs.
• Do not summarize, compress, paraphrase, add connective prose, or fill uncertain wording from context.

4. Frontmatter defaults: source: https://www.youtube.com/watch?v=<videoId>, cover: https://img.youtube.com/vi/<videoId>/maxresdefault.jpg, speaker from metadata, format matching the source, language: English, purpose: Read-aloud article, and a YAML tags: list that includes youtube.

Reference template: src/content/posts/openclaw-ted-talk.md.

Web Article URL

1. Convert with the bundled URL converter:

npx -y bun .github/skills/baoyu-url-to-markdown/scripts/main.ts <url> -o tmp/<slug>.raw.md
# Add --wait for login-required pages.

Fallback to any method that yields clean article markdown.

2. Preserve the article, not the website shell:

• Remove nav, sidebars, share widgets, cookie banners, footer links, and related-post cards.
• Preserve headings, block quotes, code blocks, and substantive figures.
• Keep original CDN image URLs as ![alt](url) in the same body position. Skip only decorative chrome images.
• Long articles stay whole unless the user asks for a summary.

3. Frontmatter defaults: source: <url>, speaker only if byline is present, format: blog post or more specific, language matching the article.

Raw Text Or Existing Markdown

• Pasted text: infer title when obvious; otherwise ask briefly. Add frontmatter and clean paragraph breaks.
• Existing markdown file: preserve body and fix only frontmatter needed for schema compliance.
• Match nearby post style: most posts rely on frontmatter title and use ## body headings, not #.

User Draft

• Preserve the user's structure and voice.
• Chinese/English mixed drafts are allowed; set language: Bilingual when both are substantial.
• Use **bold**, lists, and block quotes when they clarify the draft; do not invent bilingual translation unless asked.

Quality Gate

Before saving and before final response, check every relevant item:

• Frontmatter parses; title exists; tags is a YAML list; date is YYYY-MM-DD.
• Slug is valid and unique, or overwrite was confirmed.
• Body has blank lines between paragraphs. For non-YouTube sources, keep or add ## sections only when the source structure warrants them.
• No VTT timestamps, cue numbers, <c> tags, position tags, >, or HTML soup leaked into body.
• Web articles keep every substantive figure/diagram/screenshot from source.
• YouTube posts keep a cleaned, sentence-level transcript; no summarizing, compression, paraphrase, connective filler, or context-based gap filling.
• Long talks with Q&A are split after ---.
• No hallucinated speaker names, quotes, sources, or translations.
• Every English/bilingual post has chunks: added via english-chunks-miner before validation; Chinese-only posts omit chunks.

Anti-Patterns

• Raw VTT blocks as article body.
• Inline-array tags (tags: [a, b]).
• Files outside src/content/posts/.
• Source publish date in date.
• Guessing metadata.
• Stripping article diagrams or screenshots.
• Auto-summarizing long-form content.