hc-social - SKILL.md Agent Skill

name: hc-social description: Draft LinkedIn, X (Twitter), and Mastodon copy for a hungovercoders blog post, in voice allowed-tools: Read, Glob, Grep, Bash(ls:), Bash(mkdir:), Write argument-hint: <slug or URL — defaults to the most recent post in src/content/blog>

Draft social-media copy for a published or about-to-be-published hungovercoders blog post. Run from ~/dev/hungovercoders/site/. Outputs drafts for LinkedIn, X, and Mastodon — adapted per platform but rooted in the same voice and the same opinion beats from the post.

If $ARGUMENTS is empty, target the newest post in src/content/blog/ (excluding _drafts/). Otherwise treat the argument as a slug (2026-05-25-quick-beer-with-bento), a path, or a full URL (https://hungovercoders.com/blog/<slug>/). For a URL, parse the slug and read the local file — never scrape your own site.

Step 1 — Load context

Read in full:

~/.claude/hungovercoders/voice/datagriff-voice-guide.md — voice rules, what dataGriff is not, opinion beats, metaphor library. Section 1a (Truth in first-person) is load-bearing even for social copy — short, punchy first-person sentences are especially prone to fabrication.
~/.claude/hungovercoders/voice/facts/README.md — per-topic facts convention.
~/.claude/hungovercoders/voice/facts/<topic-slug>.md — the source of truth for first-person claims about the post's topic. Resolve the slug from the post's tags / series frontmatter, or fall back to the most prominent topic the post covers. Load only this file.

You do not need the blog template for this — the post itself is the source. You're adapting voice across platforms, not constructing one.

Truth in social copy. Social platforms compress hard — a 280-character X post or a LinkedIn opener has no room for hedging or qualifiers. That makes social copy the most exposed surface for fabricated first-person specifics. Any "I + verb" claim with a specific detail (number, duration, physical reaction, named moment) must either come from the post itself (which already passed the facts check) or be directly cited from voice/facts/<topic-slug>.md. If the post itself has unsourced specifics, flag them — drafting social copy off fabricated source material multiplies the problem. Do not invent new specifics that aren't in either source.

Step 2 — Resolve the target and extract the post's spine

Read the resolved blog post file in full. Extract:

Title (from frontmatter)
Description (from frontmatter — your one-line hook material)
Slug + canonical URL — https://hungovercoders.com/blog/<slug>/
Series link if the post references one (https://hungovercoders.com/training/<slug>)
Opener — the first 1–2 paragraphs
Honest moment — search for "I'll be honest", "first time I", "this cost me", or similar in the body
Verdict — the section with the "would I use this" or "how I actually use this" framing
The one composing demo — what the post actually built
Three or four topic-specific concrete details — Bloblang quirks, a specific config, a hook pattern, etc. These are what makes the post worth reading rather than the abstract one-liner
Themed thread — beer / dogs / food / films / geek / music — whichever the post leans on. Match it in the socials

Confirm the share image exists at public/assets/<slug>/link.png — that's the card LinkedIn and X will preview. If it's missing, surface that and stop the user before posting.

Step 3 — Draft LinkedIn

LinkedIn audience: technical professionals, hiring managers, the wider tech community. The voice guide's "what dataGriff is not — LinkedIn" rule applies: no emoji rockets, no "Excited to share that…" openers. But you can keep the self-deprecation and the worldview — that's what makes the post stand out from the LinkedIn slurry.

Shape:

Hook (1–2 lines): the post's want, or its verdict in a sharper form. State the thing you built and the take you formed.
The substance (2–4 short paragraphs, ~250–400 words total): one paragraph naming what the topic actually ships, one paragraph on the working demo, one paragraph carrying the honest moment, one paragraph with the verdict. Use line breaks between paragraphs — LinkedIn rewards whitespace.
The takeaway: a single line distilling the worldview hook from the post's verdict (small/cheap/yours, one-hungovered-person-on-a-Tuesday, etc.).
The link: Full write-up: <canonical URL> on its own line.
Tags: 3–5 hashtags, all lower-case, technical (#bento, #streaming, #dataengineering, #claudecode). Skip #hungovercoders — it's not discoverable on LinkedIn.

Length cap: ~1,300 characters including the link. Stay under it — LinkedIn truncates around 1,500.

Step 4 — Draft X (Twitter)

X audience: same tech crowd but skimming. Voice can be more pub-energy here — Welsh inflections, beer references, self-deprecation all welcome.

Draft a thread of 4–6 tweets, each under 280 characters. Shape:

Hook tweet — the want, or "I built X. Here's what I learned." Strong opening hook with a single concrete promise. End with a hook for the thread ("🧵" is banned by the voice guide; use "Thread —" or "Walking through this:" or no marker at all).
Tweet 2 — what the topic actually is, in one tight sentence.
Tweet 3 — the demo / what got built. Mention the composition ("Three things working together: X, Y, Z").
Tweet 4 — the honest moment, kept short.
Tweet 5 — the verdict, distilled. The "I'd use this for X, not Y" beat.
Tweet 6 (closer) — link to the blog: Full thing: <canonical URL> and optionally the series link if it's a launch post.

Also draft a single-tweet variant for the user who prefers not to thread: 280 chars, hook + verdict + link.

Step 5 — Draft Mastodon

Mastodon audience: technical, federated, often Linux/open-source-curious. Voice the same as X but a touch more technical, less performative. The 500-character cap gives breathing room.

Single toot, ~400–500 chars. Shape:

Opening sentence with the want
One sentence on what was built
One sentence carrying the honest moment or verdict
Link to the blog
2–3 hashtags (#bento, #dataengineering, #streaming — Mastodon search runs on hashtags more than X does)

Step 6 — Voice checks across all three drafts

Before printing, scan every draft for:

Banned LinkedIn language in the LinkedIn draft: no "Excited to share", no rockets, no "thrilled", no "passionate about"
No em-dashes as decoration — aim for zero em dashes in each draft. Short social copy has no room for the parenthetical clauses that make em dashes tempting. Replace with commas, colons, or rewritten sentences (see Section 8a of the voice guide)
No "journey" (unless literal), no "leverage" (verb, except ironic), no "robust", no "production-grade", no "seamless", no "stakeholder"
No throat-clearing phrases — "It's worth noting…", "This is important because…". Start with the thing itself.
British spellings: flavour, colour, behaviour, organisation
Voice texture — each draft should contain at least one self-deprecating or honest beat. Not a manufactured one — pull from the post's actual honest moment

Rewrite any draft that fails until it passes.

Step 7 — Report

Print all three drafts to chat in this order:

## LinkedIn

<draft>

(NNN characters)

## X — thread (N tweets)

1. <tweet 1>
2. <tweet 2>
...

## X — single-tweet variant

<draft>

(NNN characters)

## Mastodon

<draft>

(NNN characters)

---
Share image (auto-rendered preview):
public/assets/<slug>/link.png

Optionally, if $ARGUMENTS contains --save, write the full set to ~/.claude/social/<slug>-YYYY-MM-DD.md for record-keeping. Otherwise it stays in chat — these are one-shot drafts the user copies into the platforms manually.

Notes on running this

Run after the post is live on https://hungovercoders.com — i.e. after the publish branch has been merged to main and Cloudflare has deployed. The drafts hard-code the canonical URL, so running this against an unpublished branch would give the user copy that points at a 404.
The share image needs to exist locally as well (public/assets/<slug>/link.png) — that's what the platforms preview when they fetch the canonical URL.
The drafts are a starting point, not final copy. Expect to tweak before posting — the user has the lived context the model doesn't.
LinkedIn and X don't auto-detect canonical URLs the same way Slack does, so the explicit link line is non-negotiable.
Do not schedule or post automatically. This skill drafts; the user posts.