Agent skills for writing Threads posts in Korean. The same files work in both Claude Code and Codex.
The rules are measured, not guessed. They come from 4 months of official Threads API reports for @yong_____jjang (2026-04 to 2026-07) — 397 post captions and self-replies, counted line by line.
Both sides are 152 characters. Line breaks don't change length. They change reading speed.
한국어 문서: README.ko.md
| Skill | What it does |
|---|---|
threads-linebreak |
Shapes line breaks and paragraphs to the measured pattern, and ships a checker that catches violations |
threads-web-publish |
The operational procedure for driving Threads on the web with browser automation. SKILL.md holds the four hard rules and the pre-publish checklist; situational detail lives in four references/ documents |
alt-text-generator |
Writes Korean alt text for an image — front-loads the key information, transcribes text inside the image, and returns both plain text and an alt="…" attribute |
threads-brand-card |
Builds card images in the account's brand: fixed brand bar, fixed colors, and a size that matches the photo it ships alongside so Threads does not crop it |
threads-html-image |
Turns text, metrics, and screenshots into a PNG by writing standalone HTML and rendering it in a real browser — the render command, why headless Chrome hangs, and the post-render visual checklist |
SKILL.md carries only the four rules you apply every time plus the pre-publish checklist. Open the one document that matches the task at hand.
| Document | Covers |
|---|---|
references/composing.md |
Body and self-reply input, photo attachment, alt text, caret and scroll traps |
references/scheduling.md |
Scheduling, finding a scheduled post again, quote cards, why the API cannot schedule |
references/recovery.md |
Fixing a published post, pinning replies, clicks you cannot undo |
references/dm-source.md |
Turning a DM thread into post material (attributing speakers and input paths) |
One line break after a comma that ends a clause. Two line breaks (one blank line) between paragraphs.
- Not every comma — only 44% of commas land at end of line. The rest are numbers and lists; leave them inline
- Average line is 25 characters; 86% are 40 or fewer
- 76% of paragraphs are 1–2 lines. Past 3 lines, look for a place to split
- Half of all lines end with a period. Comma breaks are seasoning, not the base
Full reasoning and examples live in threads-linebreak/SKILL.md (written in Korean, since that's the language the pattern describes).
python3 threads-linebreak/scripts/check_linebreaks.py draft.txtNo dependencies (Python 3.8+). It separates hard violations from judgment calls — whether a comma joins clauses or items in a list depends on context, so the script flags candidates and leaves the decision to a human.
── 분포 ──
문단 5개 · 줄 8개 · 본문 153자(줄바꿈 제외)
줄 길이 평균 19.1자 · 40자 이하 100% (실측 기준 86%)
문단 1~2줄 비율 100% (실측 기준 76%) · 구성 [2, 2, 2, 1, 1]
── 고칠 것 없음 ──
Add --json for machine-readable output. Exits 1 when violations exist, so it drops into CI or a pre-publish hook.
git clone https://github.com/Tygb99/threads-writing-skills.git
cd threads-writing-skills
./install.shinstall.sh only creates links — it never copies or deletes:
- Claude Code → symlinks at
~/.claude/skills/<skill-name> - Codex → pointer lines appended to
~/.codex/AGENTS.md
After that, a single git pull updates both.
To do it by hand, link or copy the skill folder wherever your tool reads skills from. Both tools use the same SKILL.md format (YAML frontmatter + markdown).
MIT. If you use this, consider re-deriving the numbers from your own account — these come from one person's 397 posts, and a different voice will produce a different distribution.