How to Write a Blog Post with a Claude Code Skill
Claude Code is a command-line AI assistant from Anthropic. If you use it regularly, you can extend it with skills — reusable slash commands that give Claude a specific set of instructions to follow whenever you invoke them.
Think of a skill as a written-down procedure. Instead of explaining the same process to Claude over and over in a conversation, you write it once as a Markdown file and drop it into your project. From then on, typing /skill-name runs those instructions from the top.
The /new-blog-post skill
This site has no CMS. Every page is a plain HTML file, which keeps the site fast and simple, but publishing a post involves a handful of repetitive steps — process the hero image, write the HTML, add the index card, update the sitemap. Easy to get slightly wrong. Tedious to do manually every time.
I wrote a /new-blog-post skill to handle all of it. The skill file is a Markdown document that walks Claude through:
- Collecting the title, content brief, image, tags, and SEO metadata from me
- Deriving a clean URL slug from the title
- Downloading and optimising the hero image (WebP, resized to 1400px if needed)
- Writing the full post HTML using the site’s existing template
- Prepending the post card to
blog/index.html - Updating
sitemap.xml - Asking whether to commit and push
My side of the workflow: type /new-blog-post, answer a few questions, review the output, say yes to commit.
What this actually looks like
A skill file is just Markdown. It describes what Claude should do, in what order, with what inputs and outputs. There’s nothing to install or configure beyond dropping the file in the right folder. Skills can call shell commands, read and write files, and use any tools Claude Code already has access to.
When you invoke a skill, Claude reads the instructions and executes them step by step — asking for what it needs, making decisions where the instructions allow for it, and confirming before doing anything irreversible like committing to git.
This post is the test
I typed /new-blog-post, gave Claude a brief (essentially: “I’ve been experimenting with Claude Code skills and just wrote one to publish blog posts — this is the first real test, explain how it works”), and pointed it at an Unsplash image. Everything you’re reading now — the HTML, the optimised image, the sitemap entry, the index card on the blog listing — was produced from that one interaction.
It took about two minutes from typing the command to having a commit ready to push. The image went from a 480 KB JPEG to a 120 KB WebP. The slug, meta description, tags, and excerpt were all generated and confirmed before anything was written.
This is a small example of what’s actually useful about AI tooling: not replacing the thinking, but removing the repetitive parts so you can focus on what needs a human. In this case, the human part was having something worth saying. The rest is just process.