| name | github-page-write |
| description | Write a long-form post for Anuj's hand-rolled blog at asadani.github.io. Outputs a complete topic-slug/index.html from the single light+dark template (one design, reader toggles light/dark), with inline CSS, ready to commit. Use when Anuj wants to publish on his own site. |
github-page-write
You are drafting a long-form post for Anuj Sadani's blog at asadani.github.io (also asadani.tech). The site is hand-rolled HTML — every post lives in <topic-slug>/index.html with fully inline CSS from a single template that ships both light and dark mode (the reader toggles between them; the design is identical either way).
Step 1 — Load context (mandatory, in this order)
Read these files. Do not skip.
shared/voice-rules.mdshared/voice-samples.mdshared/pet-peeves.mdshared/topic-modes.mdshared/platform-styles.md
Then fetch the live source-of-truth from Anuj's blog repo via the gh CLI (always current, never bundled):
gh api repos/asadani/asadani.github.io/contents/CLAUDE.md --jq '.content' | base64 -d
gh api repos/asadani/asadani.github.io/contents/_template/article.html --jq '.content' | base64 -d
If gh is not available or the fetch fails, ask the user to provide the template paths or auth.
Step 2 — Theme (nothing to pick)
There is one template and it ships both light and dark mode — the reader toggles between them with the nav button; the design is identical either way. There is no per-topic theme choice and no separate "security" theme. Do not ask the user which theme to use.
Keep the theming machinery intact when filling the template: the inline <head> no-flash script, the .theme-toggle button in the nav, the end-of-body toggle wiring script, and the [data-theme="dark"] CSS block. Colors come only from CSS variables — never hardcode a color that must differ between light and dark.
Step 3 — Pick slug
From the topic, propose a kebab-case slug (e.g., "Why agentic systems fail in production" → agentic-systems-production). Confirm in one line. The final output goes to <slug>/index.html in the user's blog repo.
Step 4 — Draft the post
Voice register: measured engineer, still opinionated. Long-form: target 8-12 min reading time (~1800-2800 words). Apply topic-modes.md to adjust register for the specific topic.
Apply the github-page row from shared/platform-styles.md: technical depth 5 (full code, internals, tables — the reader may run the code and arrived deliberately), headline aggressiveness 2 (plain, accurate, SEO-clear title — not a curiosity-gap), density 5 (the canonical, densest version of the argument). Honor the precedence rules in that file — voice-rules.md/pet-peeves.md are absolute, and topic mode can raise (never lower) substance.
Structure (mirrors "Moral Surrender" / "Contextual retrieval, honestly"):
- Hero block — eyebrow label (uppercase accent), H1 title, deck (one sentence subtitle), meta line (
asadani.tech · <month> 2026 · <N> min read).
- Optional banner — accent-bg card with a one-line thesis or context note.
- Numbered sections (8-12 of them) — each begins with an uppercase section-label and a serif H2.
- Side-by-side or comparison table when there's a dichotomy.
- Plain-english explainer card for any concept a non-specialist might miss (
.plain-english, numbered .pe-num circles).
- Callout for warnings/asides (
.callout, 3px accent left border).
- Code blocks for technical content — use the documented dark code-block style with token classes (
.tok-key, .tok-val, .tok-cmt, .tok-str).
- Closer section — "The Frame to Leave On" / "The Short Version" / similar — 3-line distillation.
- Footer —
asadani.tech · <month> <year> left, ← back to index link right.
Apply all rules from shared/voice-rules.md and avoid every pattern in shared/pet-peeves.md (full paths in Step 1). Cite real numbers. First-person. Code over diagrams when technical.
Emoji: follow the Emoji Control section in shared/voice-rules.md. Default level is low — long-form blog posts should rarely contain emoji. Override only if the user explicitly asks.
Step 5 — Generate the HTML
Take the template (the raw article.html you fetched), and replace every {{PLACEHOLDER}} token with real content. Preserve ALL inline CSS and class names exactly as in the template — do not modernize, do not extract to external stylesheets.
If a section in the template doesn't apply to this post, remove the section entirely rather than leaving placeholders. If a section is needed that's not in the template, add it using the documented component classes from CLAUDE.md.
Step 6 — Pre-delivery checks
- Run the regex sweep from
shared/pet-peeves.md. Any hit → rewrite the offending paragraph.
- Confirm: no
— (em-dash) joining clauses; no rule-of-three; first-person throughout.
- Confirm: every quantitative claim has a number or is labeled anecdotal.
- Confirm: the HTML is a single complete file with
<!doctype html>, <head> (with <title>, font links, inline <style>), <body> (nav, hero, sections, footer). No external CSS files referenced.
- Confirm: style profile applied — depth 5 (full code/tables), a plain accurate title (not a curiosity-gap), and the densest long-form version, per the github-page row in
shared/platform-styles.md.
Step 7 — Save and deliver
Save to file. Get today's date in YYYY-MM-DD form. Derive a kebab-case topic slug. Save the complete HTML to:
./drafts/<YYYY-MM-DD>-<topic-slug>/github-page.html
If the file already exists (re-running on the same topic same day), suffix with -v2, -v3, etc. Create directories as needed.
Print in chat:
- The topic slug.
- Reading-time estimate.
- The full save path.
- Emoji level:
<none | low | medium | high>.
- One-line instruction: "To publish: copy
github-page.html into the asadani.github.io repo as <blog-slug>/index.html and commit."
Do not auto-commit to the blog repo.
