AI-agent skill · v1.0.0

shipcraft

A complete AI-agent skill for auditing and finishing landing pages across metadata, share cards, conversion, proof, mobile behavior, sitemaps, and soft-404 routing.

Download shipcraft · View on GitHub

The Mechanical Launch Floor for Landing Pages

shipcraft is a downloadable AI-agent skill for taking a finished landing page from “looks done” to ready to ship. It applies a mechanical launch floor across raw-HTML metadata and canonical signals, Open Graph and share-card rendering, conversion wiring, truthful proof, mobile behavior, sitemap and soft-404 routing, and silent failure modes.

It is the launch companion to scrollcraft: scrollcraft decides what a page feels like; shipcraft decides whether it works as a business asset. It also works on its own.

A floor is not a checklist item to skim. It is a concrete condition that can fail. The audit tells you which conditions it can prove; the skill makes the remaining human checks explicit.

What shipcraft covers

Surface	What the audit can assert	What remains human-owned
Head	Most metadata, canonical, Open Graph, sitemap, and structured-data consistency checks	Whether the copy says the right thing
Share card	The image exists, resolves, and matches its declared dimensions	Whether the preview captures the right moment
Conversion	Form semantics, dead anchors, and `mailto:`-only primary closes	Label consistency and a real end-to-end submission
Proof	—	Whether each claim is true, current, attributable, and cleared
Phone	Viewport setup and zoom safety	Layout, navigation, tap targets, focus, keyboard behavior, and real-device interaction
Gotchas	—	Silent browser and deployment failures that need rendered inspection

The audit deliberately reads the raw HTTP response without executing JavaScript. That is a reliable floor for link-preview bots and a useful check on a page’s launch-state HTML; it is not a claim that every search engine ignores JavaScript.

Install

Download the latest package, or install it directly for Codex:

git clone https://github.com/maraja/shipcraft.git ~/.codex/skills/shipcraft
cd ~/.codex/skills/shipcraft
npm ci

Use it before and after launch

Run the strict raw-HTML audit against a local server or a public URL:

node scripts/audit.mjs --url http://localhost:4500

Use --prelaunch only while the production canonical or sitemap cannot resolve. It makes those checks explicit warnings rather than hiding them; rerun the default strict audit after deployment.

The package also renders a reproducible 1200×630 share card from a site’s own CSS and fonts:

node scripts/sharecard.mjs \
  --css site/site.css \
  --out public/og-image.png \
  --wordmark "Your brand" \
  --tag "What this page is for"

The renderer writes the PNG and a sibling .shipcraft.json manifest so the card can be reproduced after the brand changes.

What it will not pretend to verify

shipcraft is deliberately clear about its boundary. It cannot certify that social proof is real, that a mobile navigation works under a thumb, that a form submission arrived, or that a share preview was refreshed in a platform inspector. The required report asks for those checks, their evidence, and any remaining warnings instead of silently passing them.

It also refuses keyword-density targets, misleading structured data, fake scarcity, unnecessary cookie banners, hidden keyword stuffing, and score chasing that exceeds real user benefit.

Compatibility

Node.js 20.18.1+ is required.
Share-card rendering also needs local Chrome or Chromium. Set SHIPCRAFT_CHROME when browser detection needs an override.
The package works against a local server or a live domain.

Download and source

Evidence basis

The protocol, search, and accessibility floors are grounded in primary guidance. Conversion and proof patterns are practitioner judgment and should be checked against the audience and business model of the specific page.

Raw Markdown