Create a new tech blog post on 聆.tw (琳聽智者漫談), a Traditional Chinese AI-assisted tech blog. Use when the user wants to write a new blog post, create an article, draft a tech article, or publish content on 聆.tw. Triggers on requests like "write a blog post", "create an article about X", "draft a post on 聆.tw", or "help me write about X for the blog". This skill handles the full workflow: repo setup, content creation, writing in Traditional Chinese following strict editorial guidelines, and subm...
Installs into .claude/skills of the current project.
Are you the author of 003 Name Skill A0bce514?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/tools-only-003-name-skill-a0bce514)
---
name: create-blog-post
description: >
Create a new tech blog post on 聆.tw (琳聽智者漫談), a Traditional Chinese AI-assisted tech blog.
Use when the user wants to write a new blog post, create an article, draft a tech article,
or publish content on 聆.tw. Triggers on requests like "write a blog post", "create an article about X",
"draft a post on 聆.tw", or "help me write about X for the blog".
This skill handles the full workflow: repo setup, content creation, writing in Traditional Chinese
following strict editorial guidelines, and submitting a pull request.
---
# Create Blog Post on 聆.tw
This skill guides the full workflow of creating a new tech blog post on **聆.tw** (琳聽智者漫談), from repo setup to PR submission.
## Prerequisites
- `git` CLI available
- `gh` CLI authenticated with GitHub
- Write access to `bot0419/ai-talks-content`
## Workflow
### Step 1: Clone the Repository
If the blog repo is not yet cloned:
```bash
git clone --recurse-submodules https://github.com/jim60105/blog.git
cd blog
```
If already cloned but submodules are missing:
```bash
git submodule update --init --recursive
```
If already cloned and submodules exist, just ensure they're up to date:
```bash
git pull origin master
```
### Step 2: Prepare the Submodule
Enter the content submodule and ensure it's on the latest `master`:
```bash
cd 聆.tw/content
git checkout master
git pull origin master
cd ../..
```
### Step 3: Switch to 聆.tw Mode
From the project root (`blog/`):
```bash
./switch-site.sh 聆.tw
```
This creates symlinks for `config.toml`, `content/`, `static/`, and `wrangler.jsonc` pointing to `聆.tw/`.
### Step 4: Choose Section and Prepare File
List available content sections:
```bash
ls -d 聆.tw/content/*/
```
Choose the section most related to the topic. If none fits well, use `Uncategorized`.
### Step 5: Create the Post File
Create a markdown file with a slugified filename (lowercase, hyphens, English, descriptive):
```bash
touch 聆.tw/content/<Section>/my-descriptive-slug.md
```
Naming convention: use lowercase English words separated by hyphens. The slug should describe the post content concisely.
### Step 6: Write Front Matter
Read `AGENTS.md` at the project root for the latest front matter specification. The required format:
```toml
+++
title = "文章標題(正體中文)"
description = "SEO 友善的文章描述,包含所有重要關鍵字(正體中文)"
date = "YYYY-MM-DDTHH:MM:SSZ"
updated = "YYYY-MM-DDTHH:MM:SSZ"
draft = false
[taxonomies]
tags = ["Tag1", "Tag2"]
providers = [ "Felo Search" ]
[extra]
withAI = "<https://the.ai.resource/used>"
+++
```
Rules:
- `title`: Concise, SEO-friendly, Traditional Chinese
- `description`: Contains all keywords, compelling for search results
- `date`: ISO 8601 UTC format, use current timestamp
- `updated`: ISO 8601 UTC format, use current timestamp
- `tags`: Relevant tags in the format used by existing posts
- `providers`: The provider(s) of AI assistance used in writing the article, if any
- `withAI`: Brief note about AI assistance or any urls to AI resources used. This is optional but recommended for transparency.
- **NEVER** fabricate an `iscn` field — only the user can generate this
### Step 7: Write the Blog Post Content
Read `.github/instructions/quill-sage.instructions.md` at the project root for full editorial guidelines. For quick reference, see [references/writing-guidelines.md](references/writing-guidelines.md).
Key rules:
- Write in **Traditional Chinese** (zh-TW) with full-width punctuation
- Add a space between Chinese characters and alphanumeric characters
- Use inverted pyramid structure: core conclusion first, evidence second
- Avoid bullet lists unless explicitly requested; prefer natural paragraphs
- Use `##` and `###` subheadings to organize
- Address reader as 「讀者」「大家」「各位」 or 「你」, never 「您」
- Refer to the author as 「我」, never 「我們」
- Opening paragraph states core conclusion and scope
- Closing paragraph must not use slogan-style endings
### Step 8: Add Formatting and Color Shortcodes
Review the article and enhance with:
- **Bold** (`**text**`) for emphasis keywords
- *Italic* (`*text*`) where appropriate
- Color shortcodes for pros/cons:
- Green (positive): `{{ cg(body="positive text") }}` or `{% cg() %}block text{% end %}`
- Red (negative): `{{ cr(body="negative text") }}` or `{% cr() %}block text{% end %}`
### Step 9: Add Chat Shortcodes
Use chat shortcodes to create conversational content that makes the article vivid:
```markdown
{% chat(speaker="yuna") %}
Question or statement from Yuna
{% end %}
{% chat(speaker="jim") %}
Response from Jim (author, displayed as 琳, aligned right)
{% end %}
```
Available speakers: `chatgpt`, `claude`, `gemini`, `copilot`, `felo`, `jim` (author, right-aligned), `yoruka`, or any custom name. Default/`user` gets a generic user avatar.
Design conversations that naturally introduce the topic, ask clarifying questions, or surface interesting angles. The chat format should add value, not just decorate.
### Step 10: SEO Review — Rewrite Title and Description
After completing the content, re-evaluate:
1. **Title**: Rewrite for SEO. Include the primary keyword near the front. Keep it concise but descriptive. Traditional Chinese.
2. **Description**: Rewrite to include all important keywords from the article. This text appears in search results — make it compelling and informative. ~150-160 characters ideal.
### Step 11: Rename File if Title Changed
If the title was significantly revised, rename the file to match:
```bash
mv 聆.tw/content/<Section>/old-slug.md 聆.tw/content/<Section>/new-better-slug.md
```
The slug should reflect the final title content in English.
### Step 12: Zola Build Check
Before committing, run a local build to catch any formatting or shortcode errors:
```bash
zola build
```
> Get the latest Zola binary from GitHub releases if you don't have it installed. <https://github.com/getzola/zola/releases/latest>
> Read the release to get the correct binary name with version number.
> Download the `*-x86_64-unknown-linux-musl.tar.gz` for Linux, extract it, and use the `zola` binary.
### Step 13: Create Branch, Commit, and PR
All git operations happen **inside the submodule** (`聆.tw/content/`):
```bash
cd 聆.tw/content
git checkout -b post/<slug-name>
git add <Section>/new-post-file.md
git commit --signoff --author="GitHub Copilot <bot@ChenJ.im>" -m "feat: add post <descriptive-title>
Add new blog post about <topic summary>.
Co-authored-by: GitHub Copilot <bot@ChenJ.im>"
git push origin post/<slug-name>
```
Then create the PR targeting `master` on `bot0419/ai-talks-content`:
```bash
gh pr create \
--repo bot0419/ai-talks-content \
--base master \
--head post/<slug-name> \
--title "feat: add post <descriptive-title>" \
--body "Add new blog post: <title>
<brief description of content>
---
Written with AI assistance."
```
### Step 14: Request Review
```bash
gh pr edit --repo bot0419/ai-talks-content <PR_NUMBER> --add-reviewer jim60105
```
## Terminology Mappings
When writing content, apply these Traditional Chinese mappings: create = 建立, object = 物件, queue = 佇列, stack = 堆疊, information = 資訊, invocation = 呼叫, code = 程式碼, running = 執行, library = 函式庫, schematics = 原理圖, building = 建構, Setting up = 設定, package = 套件, video = 影片, for loop = for 迴圈, class = 類別, Concurrency = 平行處理, Transaction = 交易, Transactional = 交易式, Code Snippet = 程式碼片段, Code Generation = 程式碼產生器, Any Class = 任意類別, Scalability = 延展性, Dependency Package = 相依套件, Dependency Injection = 相依性注入, Reserved Keywords = 保留字, Metadata = Metadata, Clone = 複製, Memory = 記憶體, Built-in = 內建, Global = 全域, Compatibility = 相容性, Function = 函式, Refresh = 重新整理, document = 文件, example = 範例, demo = 展示, quality = 品質, tutorial = 指南, recipes = 秘訣, byte = 位元組, bit = 位元