skill-article

${var} — Selector: <skill-name> [--brand <handle>] [--banner].

• <skill-name> → announce that skill (must exist under skills/<skill-name>/).
• empty → pick the most article-worthy skill from the last 14 days of memory/logs/: prefer a skill that shipped recently, produced verified output, or hit a milestone. If nothing qualifies, log SKILL_ARTICLE_NO_TARGET, send no notification, and exit clean.
• --brand <handle> anywhere → byline/handle for the outreach footer (default: this instance's public identity from soul/, else the repo's org).
• --banner anywhere → also generate one 16:9 title banner through the Higgsfield MCP (spends credits; see Step 5). Off by default - a blank run never spends.

Examples: "aeon-update", "rug-scan --brand @myproject --banner".

Today is ${today}. Write a launch article for one skill, modeled on the security-industry "skill announcement" format: the article sells the insight the skill encodes, not the file - and then gives the file away.

Shared preamble

1. Read memory/MEMORY.md and the last 14 days of memory/logs/ - know what this skill has actually done.
2. If a soul/ directory exists, read soul/SOUL.md and soul/STYLE.md for voice. Default voice: terse, declarative, builder-wrote-this-fast.
3. Read the target's full skills/<name>/SKILL.md. Every claim in the article must trace to a line in it.

Step 1 — Mine the track record (real numbers only)

Hunt for countable proof before writing a word:

• memory/logs/ entries under the skill's ### <name> headings: runs, OK/error statuses, notable outcomes.
• output/ files the skill produced; PRs or issues it opened (gh pr list --search).
• git log -- skills/<name>/ for ship date and iteration history.

A real stat ("84 verified findings", "31 runs, 0 false alarms") becomes the headline. If no real count exists, the headline uses the mechanism instead. Never invent a number - a fabricated stat is the one thing that kills this format.

Step 2 — Find the thesis

Answer: what does this skill see that the default workflow throws away? That gap is the thesis.

Write 3 candidate contrast pairs (two-sided sentences like "production code shows what the system does; tests show what its developers thought it did") and keep the sharpest. Everything in the article hangs on this one insight.

Step 3 — Collect war stories

1-3 concrete moments where the skill caught or produced something a naive approach would miss - pulled from memory/logs/ and output/, anonymized where needed, mechanism kept visible.

If the skill has no run history yet, write "how it plays out" scenarios and label them as scenarios. Never dress a hypothetical as a case study.

Step 4 — Write the article

Structure, in order:

1. Headline: <Skill Name>: <proof stat or mechanism claim>. Bold claim, not a setup.
2. Provenance (2 short paragraphs): what it is, where it runs, ship date, track record if real.
3. Thesis section with a punchy heading (a claim, never a label): what everyone else's workflow misses.
4. Mechanics (2-4 short paragraphs) built around the contrast pair.
5. War stories: one mini-heading per story, 2-4 sentences each, ending with the generalized pattern ("The same blind spot appears when...").
6. Reframe section: the mental-model shift the reader keeps even if they never run the skill.
7. Ship it: where it lives + one-line CTA (star the upstream repo).
8. CODENAME: <Skill Name>: the target SKILL.md embedded verbatim in a fenced block (use a 4-backtick fence - skill files contain 3-backtick fences). Do not paraphrase or trim it.
9. Outreach footer: 1-2 sentences, DMs-open style, pointing at the brand handle.

Voice rules (the anti-tells)

• Short declarative sentences. Cut every sentence that only sets up the next one.
• Banned: "revolutionary", "game-changing", "unleash", "delve", "Let's dive in".
• No parallel-structure trios, no "not X, but Y" scaffolding.
• Section headings are claims or scenes ("There is an alpha in the test folder"), never labels ("Overview").
• Verify the repo/link in the CTA exists this run before naming it.
• Never reference private forks or internal accounts - describe this instance generically and point at the upstream/public repo.

Step 5 — Title banner (only with --banner)

Skip this step entirely unless --banner was passed. When it was:

1. Check the connection first. If no mcp__higgsfield__* tool is callable, the banner is skipped, not the article: note banner: HIGGS_NOT_CONNECTED for the log, point the operator at the dashboard → MCP → Connect Higgsfield in the notify, and continue to Step 6. Same on 401/stale auth (banner: HIGGS_AUTH_STALE) or insufficient credits (banner: HIGGS_NO_CREDITS). The article never fails because the banner did.
2. Build the prompt from the thesis, not the feature list. The banner is a visual metaphor for the Step 2 contrast pair - the scene that makes the insight visible. Image models garble long text, so the only text allowed in the prompt is the skill's short name (or none); the headline lives in the article, not the pixels. Match soul/ aesthetic if it defines one; otherwise: clean, high-contrast, one focal object, no collage.
3. Generate exactly one image, --ar 16:9, following the higgsfield skill's spend rules: one generation per run, one retry at most on a transient error, never re-submit a job that already succeeded. Poll to completion with a bounded number of polls; on timeout record banner: HIGGS_FAILED and move on.
4. Capture the asset URL verbatim. Never fabricate one. Banner asset URLs may be time-limited signed URLs - flag that in the notify so the operator saves it.

Embed the result at the top of the article file, next to the alts block:

<!-- banner: <asset-url> · model: <model> · job: <id> -->

Step 6 — Deliver

1. Write the article to output/skill-articles/${today}-<skill-name>.md (shell redirection; mkdir -p first). Include an <!-- alts --> comment block at the top with 2 alternative headline + thesis-heading pairs so the operator can iterate without a rewrite.

2. Notify with ./notify -f /tmp/skill-article-notify.md (write the body under /tmp/): the headline, the thesis in one line, whether the stats are real or the headline fell back to mechanism, the banner URL when one was generated (with a save-it note if the URL is signed/expiring), and a clickable link built from the run's environment:

   ARTICLE_URL="${GITHUB_SERVER_URL:-https://github.com}/${GITHUB_REPOSITORY}/blob/main/output/skill-articles/<file>.md"
   

3. Log to memory/logs/${today}.md under a ### skill-article heading: target skill, headline, thesis pair, stats-real-or-mechanism, output path, and - when --banner was passed - a banner: line (HIGGS_OK <url> | HIGGS_NOT_CONNECTED | HIGGS_AUTH_STALE | HIGGS_NO_CREDITS | HIGGS_FAILED | skipped). Status codes: SKILL_ARTICLE_OK on success, SKILL_ARTICLE_NO_TARGET when ${var} was empty and nothing article-worthy exists, SKILL_ARTICLE_NOT_FOUND when the named skill has no SKILL.md.

Limits

• This writes the article; it does not post it. Publishing is the operator's call (or a posting skill's, explicitly chained).
• The banner is opt-in and spends real Higgsfield credits - one image per run, hard cap, and its failure never blocks the article.
• Track-record mining is only as good as memory/logs/ - a skill that runs but never logs will read as unproven, and the article will say so rather than guess.
• One skill per run. Announcing a pack is a different article; run once per skill instead.