ship-it

Ship It

A Claude Code skill that catches the boring-but-important details before you ship a screen or feature.

It doesn't check your code — it checks the product surface around that code: login, checkout, billing, empty states, 404s, admin panels, and the other screens that follow well-known patterns and get half-finished anyway because nobody double-checked the boring stuff before calling it done.

The problem

Most teams don't ship broken code. They ship incomplete screens: a login form with no "forgot password," a paywall with no restore-purchases button, a pricing page with no FAQ, a cancel-subscription flow that guilt-trips the user on the way out. None of that shows up in a test suite — it shows up later, in support tickets, one-star reviews, and App Store rejections.

This skill gives an agent a structured, opinionated catalog of these patterns — organized so it can be consulted cheaply and precisely, not dumped wholesale into every conversation.

What this checks — and what it doesn't

ship-it looks for missing or incomplete product and UX details on a screen: a password field with no password-manager support, a paywall with no restore-purchases button, that kind of thing. It is not an accessibility scanner and not a security auditor. Passing every item in a pattern's checklist doesn't mean a screen is WCAG-conformant or secure — a real audit (axe, Lighthouse, a pen test) is still a separate, complementary thing to run. Standards like WCAG, Apple's HIG, and Nielsen Norman Group's usability research get cited inside individual checklist items because they're the most specific, checkable detail available for that one item, not because this skill verifies compliance with them end to end.

Using it

Once installed as a Claude Code skill, it triggers on intent, not on the word "checklist":

"I just finished the login screen, can you check it over before I ship?" "Building a pricing page for the new plan — what usually goes on one of these?" "Is our cancel-subscription flow missing anything?"

The agent picks the section, opens only that section's index, names what the product already has versus what a Fundamental/Common pattern would predict is missing, then opens only the specific pattern file it needs to check line-by-line.

What comes back

For the login example above, a real answer looks something like this — a short list grouped by confidence, not a wall of checkboxes:

Checked against the Login pattern (Common tier — assumed to apply, since this product has user accounts).

Already there: email/password fields, password visibility toggle, forgot-password recovery.

Missing:

• Credential autofill — the password field has no autocomplete="current-password", so password managers can't offer to fill it (MDN).
• Non-revealing error states — "Incorrect password" confirms the email is registered; a generic message avoids that leak.
• Rate limiting feedback — five failed attempts just reload the same form, with no cooldown message (NN/g, Visibility of System Status).

Not flagged: SAML/SSO — this is a consumer product with no B2B customers, so it wasn't raised as a gap.

That last line is the tier system at work: a pattern, or a single item inside one, only becomes a reported gap when the product's own context calls for it, not because the checklist lists it.

How it works

Checking a screen doesn't mean loading the whole catalog. An agent works through three narrowing steps: figure out which of the four sections the screen belongs to, open that section's short index to see which patterns apply, then open only the one or two pattern files it actually needs to check line by line. A login screen never pulls in the checkout catalog.

SKILL.md              →  a pure router: steps, judgment rules, and 4 section names
  sections/*.md        →  one flat index per section: pattern name + one-liner + tier
    references/*.md     →  one file per pattern: the actual checklist, opened only when that
                            exact pattern is the one being built or reviewed

Nothing upstream repeats what's downstream. SKILL.md doesn't know what's inside a Login checklist; a section index doesn't know what "Fundamental" means beyond the word itself. This is deliberate — see ARCHITECTURE.md for the full reasoning, including the two existing Claude skills this design borrows from (impeccable, media-use) and the Anthropic Agent Skills spec it follows (progressive disclosure, <500-line SKILL.md, one-hop reference files).

Three things make this more than a flat list of checkboxes:

• Sections, not platforms. The four sections — Sidekick (mobile app), Control Room (in-product web app), Storefront (marketing website), Choreography (cross-screen interaction flows) — are just entry points. A pattern like Login or Billing that shows up in more than one section still lives in exactly one reference file, so it's written once and never drifts into two contradictory versions.
• Tiers, not a flat "does this apply?" Every pattern is tagged Fundamental, Common, or Conditional — a fixed prior for how surprised the agent should be if it's missing, so "use judgment" has an actual anchor instead of drifting toward whichever answer is less work. Security and Privacy are Fundamental (assume they exist; excluding them needs a real reason). A payment flow is Common (assume it exists unless the product is free or sales-led). A shopping cart is Conditional (assume it doesn't exist until something multi-item is confirmed) — payment outranks cart because more products charge money than sell multiple items at once.
• Judgment calls are explicit, not implied. SKILL.md states the failure modes directly: don't nag a B2B tool about a Cart it doesn't need, but don't wave away a Privacy Policy just because the product is small. Once a person has answered a judgment call, the skill doesn't ask again.

Status

All 59 indexed patterns across the four sections are written and pass scripts/validate.py. The one known gap: the Web App source note never ingested its own Login pattern, so Control Room resolves Login through Mobile/Website's shared file rather than having its own row yet — see sections/control-room.md.

Billing and Login were written by hand first and set the quality bar; the remaining 57 were generated from generation/PROMPT.md against that same bar (real citations only when confident they resolve to something real — WCAG 2.2, Apple's HIG/App Store Review Guidelines, Material Design, MDN, GOV.UK Design System, and Nielsen Norman Group's usability research), then spot-checked rather than trusted blindly. What's left is deepening citations further over time — Baymard Institute's e-commerce research is a natural next source for Cart/Checkout/Pricing specifically — not filling gaps.

Project structure

SKILL.md                     the router — read this first
AGENTS.md                    repo layout, terminology, and rules for anyone extending this
ARCHITECTURE.md              design rationale: why one skill, why tiers, why these sections
CONTRIBUTING.md              how to propose a pattern or report a bad citation
sections/
  sidekick.md                mobile app pattern index
  control-room.md            web app pattern in