Claude Skills — ojitha.github.io

Project-scoped Claude skills for this repo. Each skill lives under .claude/skills/<skill-name>/ and is auto-discovered by Claude Code when run from inside the repo. They are not active in any other project.

Index

  • blog-post-excerpt — generate an inline SVG thumbnail and ~100-word excerpt for a Jekyll post and inject both into the front matter.
  • blog-post-linkedin — generate a Google-SEO-friendly LinkedIn post (≤ 50-word prose, canonical URL, comprehensive hashtag set) for a Jekyll post — printed in chat, no file output.

blog-post-excerpt

Generate an embedded SVG illustration in a chosen drawing style plus a short HTML summary, and inject both into the Jekyll front matter excerpt: field of a post under _posts/.

Location

.claude/skills/blog-post-excerpt/
├── SKILL.md
└── references/
    └── style-guide.md

What it does

For a single blog post in _posts/, the skill:

  1. Reads the markdown file end-to-end.
  2. Picks 3–6 technical jargon terms that genuinely characterise the post (framework names, algorithms, hardware, concepts — not generic words).
  3. Builds a 150 × 150 px coloured SVG in the requested drawing style, with the chosen jargon rendered as readable <text> inside the artwork.
  4. Writes a ~100-word HTML excerpt (a single <p> block, with 2–4 of the same jargon terms wrapped in <b>).
  5. Replaces <image> (or the older typo <imgage>) with the SVG and TLDR with the excerpt — directly inside the existing one-line excerpt: value in the YAML front matter.
  6. Re-reads the front matter to verify the YAML still parses cleanly.

The body of the post is never touched.

Jekyll front-matter prerequisite

Before the skill runs, the target post must already have an excerpt: placeholder line in its YAML front matter. The standard placeholder shape used in this repo is:

---
layout: post
title:  <Your post title>
date:   YYYY-MM-DD
categories: [<category>]
toc: true
mermaid: true
maths: true
typora-root-url: /Users/ojitha/GitHub/ojitha.github.io
typora-copy-images-to: ../../blog/assets/images/${filename}
excerpt: '<div class="image-text-container"><div class="image-column"><imgage></div><div class="text-column">TLDR</div></div>'
---

The skill looks for either spelling — <imgage> (the historical typo found in older posts) or the corrected <image> — and treats them identically. After the skill runs, that one line is rewritten in place: the <imgage>/<image> token is swapped for the inline SVG, and TLDR is swapped for the ~100-word HTML excerpt. Every other field in the front matter is preserved exactly as it was.

If a post has no excerpt: line at all, the skill inserts the line above (with the <imgage> placeholder) immediately before the closing ---, then performs the two replacements. So you can use the skill on a brand-new post that was never set up with a placeholder, but adding the placeholder yourself in your post template is cleaner — that way every new post is “skill-ready” from the moment you create it.

Required inputs

Parameter Example Notes
Markdown filename 2026-05-05-Gemma4.md Bare name or full path. The skill resolves to /Users/ojitha/GitHub/ojitha.github.io/_posts/<filename>.
Drawing style oil One of pencil sketch, oil, impressionist, watercolor, charcoal, pop art, or any other style — the skill will improvise within its principles.

If either is missing, the skill asks once and then proceeds.

How to invoke

In Claude Code, just describe what you want — supply both inputs in one message. Any of these phrasings will trigger the skill:

  • Run blog-post-excerpt on 2026-05-05-Gemma4.md in oil style.
  • Add a TLDR to 2026-04-08-ObsidianMethod.md as a watercolor.
  • Create an excerpt for 2025-10-31-Functional-Programming-Abstractions-in-Scala.md, pencil sketch.
  • Generate the front-matter excerpt for 2026-03-07-ContainerRocm.md in pop art style.
  • Draw a thumbnail for 2025-11-07-SparkDataset.md, impressionist.

You don’t have to say “skill” or use the exact name — pairing a _posts/*.md filename with a drawing style is enough.

Drawing styles

Each style has a complete worked SVG example in references/style-guide.md. Summary:

Style Visual feel Palette
pencil sketch Thin grey strokes, light cross-hatching, muted pastel washes Cream background, soft pastels
oil Thick saturated strokes, no fine detail, no outlines, bright highlight Deep umber/indigo background, rich reds/ochres/viridian, cream text
impressionist Many small coloured dabs, no outlines, suggestion over definition Light pastel-bright
watercolor Translucent pools with feGaussianBlur bleeding edges Soft blues, peaches, greens on off-white paper
charcoal Dark rough strokes with smudged shading, one faint colour accent Sketchpad cream + near-black + single warm accent
pop art Bold flat colours, thick black outlines, optional benday dots Bright primaries + black + cream

Any unlisted style works too — the skill leans on the closest listed one and adjusts palette and stroke weight.

What changes in the post

Only line 11-ish of the front matter — the existing excerpt: line — is rewritten. Everything else (layout, title, date, categories, mermaid/maths/toc flags, typora roots, the entire body) is left exactly as it was. Re-running the skill on the same file with a different style cleanly overwrites the previous SVG and excerpt because the placeholder structure is identical.

If a post has no excerpt: line at all, the skill inserts one immediately before the closing --- of the front matter using the standard template.

Verification (the skill always runs this before reporting done)

After editing, the skill re-reads the first ~20 lines of the file and confirms:

  1. Front matter still opens and closes with ---.
  2. The excerpt: line starts with '<div class="image-text-container"> and ends with </div></div>'.
  3. No unescaped single quotes inside the YAML value (apostrophes are escaped as '', or rephrased out).
  4. The SVG block contains <svg … </svg> and the chosen jargon words appear as <text> content.
  5. The text-column prose is roughly 100 words (counted, not estimated).

Then it reports the path to the updated post and the jargon words it chose.

YAML safety, in case you edit by hand

The full excerpt value lives inside YAML single quotes on one physical line. Anything you add inside it must obey:

  • All HTML/SVG attributes use double quotes (fill="#abc", never fill='#abc').
  • Any literal single quote in the prose must be doubled (it''s, not it's).
  • Don’t switch the outer quoting to double quotes — other tooling in the repo expects single-quoted excerpts.
  • No literal newlines inside the value.

Example output

A worked example lives in _posts/2026-05-05-Gemma4.md (oil style, jargon: Gemma 4, vLLM, ROCm, Ryzen AI, NPU). Re-read line 11 of that file to see the exact YAML shape.

blog-post-linkedin

Generate a Google-SEO-friendly LinkedIn post that promotes a single Jekyll blog post under _posts/. The output is plain text printed in the chat response — copy-paste it directly into LinkedIn. The prose is capped at 50 words, and the hashtag set is comprehensive (typically 6–10 tags) so it carries the SEO breadth that the short prose cannot.

Location

.claude/skills/blog-post-linkedin/
└── SKILL.md

What it does

For a single blog post in _posts/, the skill:

  1. Reads the markdown file end-to-end and extracts the title, category, lede, and recurring technical entities.
  2. Derives the canonical published URL from the filename and front matter (formula below).
  3. Picks 5–8 SEO entities — the proper nouns real readers Google for (framework names, model names, hardware, tools).
  4. Drafts a tight LinkedIn post:
    • Prose capped at 50 words total (hook + body combined). Numbers like 128K, hyphenated tokens like decoder-only, and proper-noun model names are each counted as one word. The URL line and the hashtag line do not count toward the 50.
    • The published URL on its own line.
    • All relevant hashtags (typically 6–10) on the final line — a mix of broad-category, entity-specific, and optional audience tags for maximum discoverability.
  5. Prints the post directly in the chat response inside a fenced code block, so the user can copy-paste straight into LinkedIn.

The original blog post is never modified, and no file is written to disk.

Required input

Parameter Example Notes
Markdown filename 2026-05-05-Gemma4.md Bare name or full path. The skill resolves to /Users/ojitha/GitHub/ojitha.github.io/_posts/<filename>.

URL derivation rule

The published URL is built deterministically from three pieces of the source post:

  • Category — lowercase the first entry in the front-matter categories: list.
  • Date — split the YYYY-MM-DD from the filename (or front-matter date:) into YYYY/MM/DD.
  • Slug — the filename with the leading YYYY-MM-DD- prefix and trailing .md removed.

Combined as:

https://ojitha.github.io/<category>/<YYYY>/<MM>/<DD>/<slug>.html

Worked example. 2026-05-05-Gemma4.md with categories: [AI] → https://ojitha.github.io/ai/2026/05/05/Gemma4.html

Edge cases handled by the skill:

  • Multiple categories → uses the first.
  • Spaces in the slug (older posts) → URL-encodes them as %20.
  • Front-matter date: disagreeing with the filename date → prefers the filename date (that is what Jekyll uses).

How to invoke

In Claude Code, describe what you want and supply the filename:

  • Run blog-post-linkedin on 2026-05-05-Gemma4.md.
  • Write LinkedIn copy for 2026-04-08-ObsidianMethod.md.
  • Make a LinkedIn announcement for 2025-10-31-Functional-Programming-Abstractions-in-Scala.md.
  • Promote 2026-03-07-ContainerRocm.md on LinkedIn.
  • Share 2025-11-07-SparkDataset.md on LinkedIn — SEO-friendly.

You don’t need to say “skill” — pairing a _posts/*.md filename with “LinkedIn” is enough to trigger.

Output — in-chat text only

The skill prints the post directly in the chat response inside a fenced code block, like:

```
<HOOK + body, ≤ 50 words>

Read more:
https://ojitha.github.io/<category>/<YYYY>/<MM>/<DD>/<slug>.html

#Tag1 #Tag2 #Tag3 #Tag4 #Tag5 #Tag6 ...
```

After the code block, the skill summarises the prose word count, the derived URL, and the hashtag list, so you can sanity-check before pasting.

No file is created. No directory is written. Re-running the skill simply produces a fresh draft in the next response — there is nothing to clean up on disk.

Writing rules the skill follows

  • Plain text only. LinkedIn’s main feed does not render markdown — no **bold**, no # headers, no [label](url). Bare URLs become clickable on their own.
  • Short paragraphs with blank lines between them — the LinkedIn idiom, easier to scan on mobile.
  • Concrete entities, repeated naturally — 2–3 mentions across the post is what Google indexes.
  • Numbers over adjectives — “128K context window” beats “huge context window”.
  • No emojis unless the source blog post uses them — match the blog’s serious technical voice.
  • No clickbait — undermines technical credibility.
  • First-person is fine because you own the blog.

Hashtag selection — comprehensive coverage

Because the prose is capped at 50 words, hashtags are the primary SEO surface area. The skill includes every relevant tag, mixing three scopes:

  • Broad category tags (1–3) — match the post’s overall domain: #MachineLearning, #GenerativeAI, #DataEngineering, #Scala, #AWS, #Kubernetes, #Python, #FunctionalProgramming, #CloudComputing, #DevOps, #BigData.
  • Entity-specific tags (3–7) — one per major SEO entity from the post: #Gemma4, #vLLM, #ROCm, #RyzenAI, #NPU, #Docker, #PyTorch, #ApacheSpark, #PySpark, #Terraform, #ElasticSearch, #LangGraph, #AWSCDK, #Hadoop, #HDFS.
  • Audience tags (0–2, optional) — when the post targets a specific community: #OpenSource, #LLM, #EdgeAI, #MLOps, #DataScience.

Format rules:

  • PascalCase for multi-word tags (#MachineLearning, not #machinelearning).
  • For single-word tools, retain the canonical spelling (#vLLM keeps the lowercase v that the project itself uses; #ROCm keeps the lowercase m AMD itself uses).
  • All hashtags on a single final line, single-spaced, no duplicates, no punctuation.
  • Aim for 6–10 hashtags total — enough for comprehensive discoverability without triggering LinkedIn’s over-tagging penalty.

Verification (the skill always runs this before reporting done)

  1. Prose word count is ≤ 50 — counted, not estimated.
  2. The published URL is correct — re-derived from the filename and string-compared.
  3. All relevant hashtags are present — at least one broad tag and 3+ entity-specific tags. Total 6–10.
  4. No markdown leaked through (no **, [, ](, # ).
  5. The output is a code block in the chat response. No file was written.

Then it reports the prose word count (e.g. “47/50 words”), the derived URL, and the hashtags chosen.

Adding a new skill to this project

  1. Create .claude/skills/<your-skill-name>/SKILL.md with the standard YAML frontmatter (name, description).
  2. Optionally add references/, scripts/, or assets/ subfolders — Claude reads them on demand.
  3. Add a section to this file describing the skill’s purpose, inputs, invocation phrases, and any verification it runs.
  4. Project-scoped skills only activate when Claude Code is launched from inside this repo.