conversion-review · git:20261104.4063c88 · 2026-11-04 · sha256 fed399376a10d4c0
conversion-review git:20261104.4063c88A
Immutable. This exact content never changes and is served at /api/v1/blob/fed399376a10d4c0.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_xxyguvk3cdh75tca/fed399376a10d4c070bc12313d727eb2cb261a7dd27e77b9b16f368394e35d87.json
Then /plugin install conversion-review@markdownregistry-xxyguvk3cdh75tca in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: conversion-review
description: >-
Audit a page, screen, screenshot, artifact or URL against conversion
heuristics: value proposition on the first screen, one primary action per
screen, proof, sign-up and checkout friction, trust, pricing clarity and
mobile. Returns prioritized findings, each with the fix, a severity and the
metric or event that would prove it. Trigger on "conversion review", "why
does nobody sign up", "audit this landing page", "review this pricing
page", "will this page sell", "check the funnel on this screen", or before
shipping any page whose job is to get a sign-up or a payment.
---
# Conversion review
**Product name.** `{{app}}` here and in the references stands for the user's product. Take its name from the project (README title, package.json, app.json, Info.plist) or ask once, then write every example with that name and the product's own screens.
A page that sells answers three questions in the first five seconds: what is
this, is it for me, what do I do next. Everything below checks whether the
page answers them, and then whether anything between "yes" and "paid" gets in
the way.
## Inputs
Accept any of: a live URL (fetch it, and fetch it again with a phone user
agent), a screenshot or several, a published artifact, or the source files of
the page. Ask one question if it is missing: **what is the one action this
page exists to get** (sign up, start trial, buy Pro, install the app). Without
it every finding is a guess.
If the project has analytics, read the event names before writing findings,
so the "prove it" column uses events that exist or names the one to add.
## The pass
Walk the page top to bottom once, then the path to payment once. Score each
heuristic; details and examples for each are in
`references/heuristics.md`.
1. **First screen (above the fold, at 390x844 and at 1440x900).** A headline
that states the outcome in the visitor's words, a subline that says for
whom, one visible primary action, one piece of proof. A logo, a slogan and
a carousel is not a value proposition.
2. **One primary action per screen.** Exactly one filled button per viewport.
Secondary actions are links or outlined. Two equal buttons split the click.
3. **Proof.** Specific, checkable, near the claim: a public track record
({{app}}: every past prediction next to the actual result, including the
misses), numbers with a date, named users, press. Unverifiable superlatives
count as zero proof.
4. **Friction to sign-up.** Count fields, screens and decisions between the
primary action and value. Social sign-in offered, email verification not
blocking the first use, no credit card for a free tier, no account wall in
front of the thing the page promised.
5. **Friction to checkout.** Price visible before the checkout form, the
store's native sheet on mobile (Apple Pay, StoreKit), no re-entry of data
already known, errors inline and recoverable.
6. **Trust.** Who is behind it, how to cancel, refund terms, privacy in one
sentence, working contact. Broken images, lorem ipsum, a stale copyright
year and a 2019 changelog all cost trust.
7. **Pricing clarity.** What the free tier gets, what Pro adds, the total
charged and the billing period in the same place. See the
`paywall-and-pricing` skill for the full gate and pricing checks.
8. **Mobile.** Thumb-reachable primary action, 44pt targets, no horizontal
scroll, text readable without zoom, the page loads its first screen fast
on a mid-range phone (LCP under 2.5 s).
9. **Speed and stability.** Layout shift when fonts or images load, spinners
in place of content, a hero that waits on JavaScript.
10. **Copy.** Jargon on a page for newcomers (see the `ux-copy` skill), a
call to action that describes the click ("Continue") instead of the
outcome ("See this weekend's prediction").
## Severity
- **Blocker**: stops a willing visitor from converting (broken checkout, no
price anywhere, primary action off-screen on mobile, an account wall before
any value).
- **Major**: measurably lowers conversion for most visitors (no proof, two
competing primary actions, vague headline, card required for a free tier).
- **Minor**: polish that compounds (weak microcopy, one extra field, slow
secondary image).
Rank by severity, then by how many visitors hit it (first screen beats
footer), then by effort (cheap fixes first within a tier).
## Output
One table, then the notes.
| # | Where | Finding | Fix | Severity | Proves it |
|---|-------|---------|-----|----------|-----------|
| 1 | Hero, mobile | Primary action below the fold at 390x844 | Move "See this weekend's prediction" above the hero image; image becomes background | Blocker | `landing_cta_tap / landing_view` up; scroll-depth before first tap down |
Rules for the table:
- **Fix** is concrete enough to implement without a meeting: the new copy,
the element to move, the field to delete.
- **Proves it** names an event or ratio. Use the project's taxonomy; when an
event does not exist, write it as `new: <event_name>` and say what fires
it. Every event plan covers entry, success, failure and drop-out.
- Do not list more than 15 findings. If there are more, the page needs the
`landing-builder` skill, not a review.
After the table: **Top three to ship first** (one line each), then **What
this review could not see** (logged-in states, real traffic mix, load under
real network), then **Open questions**.
## Rules
- Honest conversion only. Never recommend a pattern listed as forbidden in
`paywall-and-pricing` (fake scarcity, confirm-shaming, hidden cancellation
and the rest), even when it would move the number.
- Review what is on the page, not what the team meant. If a claim needs
context to be true, the page needs the context.
- A hypothesis is not a finding. "Users might prefer a video" goes to the A/B
plan in `landing-builder`, not into the table.