Skip to content

Checks

Landing page critic

Checks a page for message match, clarity, proof and friction, weighted by page type, and ranks the fixes with rewritten lines.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick landing-page-critic.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Heuristic review in conversion research (ResearchXL), 3 copy tests
Needs
The page · The page's one goal
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
2.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedCreative
SkillCreative and conversion

Landing page critic

Example output

Review a landing page
CheckSeverityResult
The page makes the ad's offerCriticalFail
Proof beside the main claimMajorFail
One primary actionMajorPass

Checks a page for message match, clarity, proof and friction, weighted by page type, and ranks the fixes with rewritten lines.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
None. It's instructions only.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

landing-page-critic/SKILL.md

In the folder

  • landing-page-critic/SKILL.md
  • landing-page-critic/references/checks.md
---name: landing-page-criticdescription: >-  Reviews one B2B SaaS page for message match, clarity, proof and friction, with pass, fail or not run per check,  severities weighted by the page type (homepage, pricing, comparison, campaign landing page, feature), fixes  ranked by impact with rewritten lines, and a PASS, PASS WITH NOTES, FAIL or BLOCKED verdict. Use it before a page  goes live, when a page converts poorly, or to tear down a competitor's page. Not for a full funnel diagnosis with  analytics (use conversion-research), for checking ads (use ad-creative-critic) or for tone alone (use  voice-check).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads the live page and takes  screenshots if the agent can browse; otherwise works from screenshots or pasted copy and names the checks it  couldn't run.metadata:  author: SHAKEUP  version: "2.0.0"  stage: check  plugin: creative-conversion  homepage: https://www.shakeuphq.com/skills/landing-page-critic--- # Landing page critic Checks a page the way its buyer meets it, straight from the ad or the search, and ranks what to fix first. ## Purpose - One job: run a fixed set of checks on one page and rank the failures by impact.- Stage: check. It ends in a verdict with severities.- It judges against the page's buyer, its traffic source and its goal, never against taste. It never marks a check it couldn't run as a pass, and never fixes copy by adding a claim the profile doesn't support. ## Use when / not | Situation | Use || --- | --- || A page is about to go live, or was just rebuilt | this skill || A page converts poorly and you want the likely causes on the page | this skill, then `conversion-research` if the cause isn't on the page || Tearing down a competitor's page | this skill (P4 is `not run`: their sources aren't visible) || Diagnosing a whole funnel with analytics and device data | `conversion-research` (it runs this skill as its heuristic pass) || Checking ads before they run | `ad-creative-critic` || Only the wording and tone | `voice-check` || Testing whether buyers understand the message | `message-test` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The page | yes | the live URL, read in a browser at desktop and phone width; screenshots the reader uploads (first screen on desktop and phone, then the full page); the copy pasted in order, with button labels and form fields | ask once; with nothing readable, `BLOCKED` || The traffic source | no (yes for a campaign page) | the ad (headline, text, image, offer), the search terms or the email that sends people here; one per main source | M checks are `not run`; on a campaign page coverage is `partial` and the verdict can't be `PASS` || The page's goal | yes | the reader's note; inferred from the primary button and confirmed | ask once; if no answer, infer and say so || Buyer phrases | no | a `message-mining` handoff; buyer quotes the reader pastes | use `### Buying group` objections only || Analytics for the page | no | a connected analytics tool; an export (conversion rate by device, form starts against submits, scroll depth) | judge from the page alone; nothing marked `confirmed by data` | What each form of the page allows: | You have | Not run || --- | --- || Live page or desktop and phone screenshots | nothing || Desktop screenshot only | F5 || Pasted copy only | C7, C9, M4, F5, F6, F7 | ## Profile read From `company-profile.md`: `## Company` (category, business model), `## Buyers` and `### Buying group` (who the page is for and each role's objection), `## Positioning` and `## Messaging house` (what the page should say), `## Voice` (for rewrites), `## Claims and proof` and `### Never say` (P4), `## Competitors` (mention policy, for comparison pages), and every `## Learnings` entry for `landing-page-critic` with status `noted`, `fact` or `accepted`. No profile: ask in one message who the page is for, what the product does, and which figures on the page the company can source. Say that `company-profile` saves this next time. Without claims, every first-party number on the page is `unsourced` in P4. ## Steps 1. **Apply learnings.** An accepted learning that changes a weight or a severity overrides the defaults here. Name it when it fires. 2. **Record what you can see.** Note which forms of the page you have (table above) and list the checks that are `not run` and why before you start. 3. **Set the page type, the goal and the weights.** Page type, first match wins:    | Signal | Page type |   | --- | --- |   | Built for a campaign: the reader says an ad or email sends traffic here, or a campaign URL (such as /lp/) with one goal. A full site menu on it is an F3 fail, not a different type | campaign landing page |   | The URL or H1 says vs, alternative or compare | comparison |   | Plans with prices, or a /pricing URL | pricing |   | The domain root | homepage |   | One capability, a /product or /features URL | feature |    Area weights by page type:    | Page type | Match | Clarity | Proof | Friction | Its own checks | Why |   | --- | --- | --- | --- | --- | --- | --- |   | Campaign landing page | 3 | 2 | 2 | 3 | none (M, F2, F3 carry it) | Each visit was paid for, so a broken promise or a hard next step wastes the click |   | Homepage | 1 | 3 | 2 | 1 | H, weight 3 | Visitors arrive from everywhere; the page must make the category and the path clear |   | Pricing | 1 | 3 | 2 | 3 | PR, weight 3 | Buyers come to check fit and cost; unclear plans and hard steps lose them at the last step |   | Comparison | 2 | 2 | 3 | 1 | CP, weight 3 | Readers are skeptical and comparing; proof and fairness decide it |   | Feature | 2 | 3 | 2 | 1 | FE, weight 2 | One capability must be clear and tied to a result |    Findability checks (T) run only on pages meant for search or AI answers, at weight 1. 4. **Read the first screen only.** Before scrolling, write the 5-second read as a stranger would: who it's for, what it does, what they get, what to do next (C1). Then run M1 to M4 against each traffic source. Most visitors decide here, so these findings come first in a tie. 5. **Run the rest of the checks.** Work through `references/checks.md` for every area and the page type's own group. Record pass, fail or not run for each, and quote the page for every fail. Use buyer phrases from `message-mining` to judge C2 and P5: a headline in the company's words when buyers use different ones fails C2's "could no one else say it" more often. 6. **Check the claims (P4).** Run `claim-check` on the page copy. Without it: list every number, comparison and quote, and match each to `## Claims and proof`; anything unmatched is `unsourced`, anything on `### Never say` is `forbidden`. Prices, plan limits and product facts on the company's own page are its offer, not claims; results, customer counts, ratings and comparisons are claims. 7. **Rank by impact.** Score each fail: severity points (critical 10, major 3, minor 1) times the area weight. The gap between 10 and 3 makes a critical outrank any major on any page type. Ties: first screen first, then the cheaper fix (a copy change before a layout change before a rebuild). If analytics back a finding (the phone conversion rate is far below desktop on a page that fails F5), mark it `confirmed by data`; never raise a severity from someone else's benchmark. 8. **Write the fixes.** For every critical and major: the change, and for copy, the rewritten line in the company's voice. Build rewrites from the source's promise, buyer phrases and `## Claims and proof` only. Recommend a new section only to fix a failed check (an objection with no answer); page length alone is not a finding. Then run `voice-check` on the rewrites and `claim-check` on any rewrite with a number or a comparison. Without those skills: no dashes, nothing from `### Words to avoid`, the 3 copy tests, and every number traced to `## Claims and proof`. Fix what fails and recheck, at most 2 rounds; label any rewrite still failing. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | Nothing of the page could be read (no live page, screenshot or copy), or the first screen is missing || `FAIL` | 1 or more critical findings, or 3 or more majors (each loses some buyers, and they add up along the page) || `PASS WITH NOTES` | Fewer than 3 majors, or minors only; or no fails but a weight-3 area for this page type was `not run` || `PASS` | No fails, and every check in the weight-3 areas ran | Named stop: `BLOCKED` asks for screenshots or the pasted copy. No other stops: an unknown goal is inferred from the primary button and stated; an unknown page type is inferred and stated. ## Learn Append to `## Learnings` in the profile: - The reader rejects a finding ("our buyers expect a long form, sales needs these fields"): a `proposal` to change that check's severity for this company, with their reason.- A fix ships and a result comes back (conversion rate before and after from the reader's analytics, or a test result): a `fact` with both rates, the dates and the source.- 2 or more results point the same way (every proof fix on comparison pages lifted demos): a `proposal` to change that area's weight for this company. Example: `- 2026-10-24 | landing-page-critic | fact | Moving the customer quote beside the demo form lifted form completion from 2.1% to 2.9% | Why: GA4, demo page, Aug against Sep 2026 | Status: noted` ## Output ```Verdict: <PASS | PASS WITH NOTES | FAIL | BLOCKED> · Coverage: <complete | partial (what's missing) | blocked>Page type: <type> (<given | inferred from ...>) · Goal: <action> · Traffic source: <source or not given>Seen: <live page | desktop screenshot | phone screenshot | copy> · Not run: <ids and why>Counts: <n> critical, <n> major, <n> minor · Learnings applied: <list or none> 5-second read: For: <...> | Does: <...> | Get: <...> | Next: <...> (<n> of 4) | Rank | Check | Area | Result | Severity | Evidence (quoted) | Fix |(fails ranked by impact, then passes, then not run) Top 3 fixes, in order:1. <change> | Rewrite: "<line>" | First because: <severity x weight, first screen>2. ...3. ... Keep: <up to 3 things that work, quoted, so a rewrite doesn't lose them>Rewrites checked: voice-check <verdict>, claim-check <verdict> (<or: checked inline>)``` Then the handoff block. ## Hand off ```handoffskill: landing-page-criticversion: 2.0.0verdict: FAILcoverage: partial (no phone screenshot)for: <page URL or name>outputs:  page_type: campaign landing page  critical: 2  major: 3  top_fix: <one line>  failed: [M2, C5, F1]  not_run: [F5]next: [conversion-research, message-test]open: [phone screenshot for F5]``` `conversion-research` reads `failed` as its heuristic pass for this page. After the fixes ship, `message-test` can pre-test the new first screen. ## Failure modes | Problem | Do this || --- | --- || The page is behind a login, blocked, or drawn by JavaScript so a fetch shows almost no text | Ask for screenshots and the copy; on a search page, the empty HTML is itself a T2 fail || Several traffic sources make different promises | Run M checks per source; if they conflict, flag that one page can't match them all (split the page or align the ads) as a major || The reader asks for opinions on colours or brand style | Answer only through a check with a rule (C7 hierarchy, M4 continuity); taste is not a finding || A competitor's page | Run everything except P4 and CP checks that need their sources; compare their first screen with the reader's || The page has 2 goals (book a demo and start a trial) | Ask which one counts; until then, C4 fails as major and the review uses the one higher on the page || The reader asks for more sections | Add a section only where a failed check needs one; say which check | ## Influenced by - CXL's conversion research framework (ResearchXL), whose heuristic pass reviews each page against fixed questions about relevance, clarity, friction and distraction, and treats the result as an informed opinion to confirm with data. We turned the questions into checks with severities, weighted by page type.- Harry Dry's 3 copy tests (can I picture it, can it be proved wrong, could no one else say it), used in C2: https://www.demandcurve.com/newsletters/growth-newsletter-202- SHAKEUP's landing page teardown method: message first, then the next step, then findability and weight, with the first screen read before anything else.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

We turned an open-ended review into fixed checks with severities, weighted by page type, so 2 reviewers reach the same verdict. The first screen is read before anything else, every fix is ranked by impact, and rewrites can't add a claim you can't back.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Note what it can see

    Pasted copy can't show hierarchy, phone layout or speed, so those checks are marked not run before it starts. A check it couldn't run is never a pass.

  2. 02

    Weigh by page type

    A campaign page weighs message match and friction most, a pricing page weighs plan clarity, a comparison page weighs proof and fairness.

  3. 03

    Read the first screen alone

    It writes what a stranger gets in 5 seconds: who it's for, what it does, what they get and what to do. Then it checks the first screen against the ad's promise and offer.

  4. 04

    Run every check and quote the page

    Each check passes, fails or doesn't run, with the exact line or element as evidence. Every figure on the page goes through the claim check.

  5. 05

    Rank by impact

    Severity times the area's weight for this page type. A critical always outranks a major, and first-screen fixes win ties.

  6. 06

    Write fixes that pass your checks

    Each critical and major fix comes with a rewritten line built from your messaging and approved claims, run through the voice and claim checks.

What you need

  • Required

    The page

    The live page, read at desktop and phone width if your agent can browse → Screenshots you upload, first screen on desktop and phone, then the full page → The copy pasted in order, with button labels and form fields

  • Optional

    The ad, search or email that sends people to the page

    The ad's headline, text, image and offer → The search terms or the email

  • Required

    The page's one goal

    Your note → Inferred from the main button, then confirmed with you

  • Optional

    Your buyers' own words

    A message mining handoff → Buyer quotes you paste

  • Optional

    Analytics for the page

    A connected analytics tool → An export by device, with form starts and submits

What it can return

  • FAIL

    Any critical finding, or 3 or more majors.

  • PASS WITH NOTES

    Fewer than 3 majors, or only minor findings. Also when a heavily weighted area couldn't be checked.

  • PASS

    Nothing failed and every heavily weighted check ran.

  • BLOCKED

    It can't read the page or its first screen.

Where it stops for you

  • When it can't read the page, it asks for screenshots or the copy

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells security questionnaire software. The page is a campaign landing page for a paid social ad that promises same-day answers and a 3-minute tour.

Verdict

FAIL

Coverage: partial (copy only, so hierarchy, phone layout and speed weren't checked)

Page type: campaign landing page

5-second read

1 of 4. A stranger can't tell who it's for, what it does or what they get, only that the button asks for a demo.

RankCheckSeverityEvidenceFix
1The page makes the ad's offerCriticalAd: a 3-minute tour. Page: "Request a demo"Lead with the tour, and keep the demo as a second step
2The first screen keeps the promiseCriticalHeadline: "The trust platform for modern teams""Answer security questionnaires the same day."
35-second readCritical1 of 4 questions answeredName the buyer and the result in the line under the headline
4Commitment fits the trafficMajorCold paid social traffic gets only a sales callOffer the tour first
5Form lengthMajor8 fields, including budget and phoneCut to email, company and role
6Proof is specificMajor"Great product!" with no role or resultUse a cleared quote with a role and a result
Top 3 fixes, in order

make the tour the main action, put the ad's promise in the headline, then cut the form to 3 fields. The first 2 are critical and sit on the first screen.

Keep

the product screenshot halfway down, which shows a questionnaire being answered.

The handoff block the next skill reads
skill: landing-page-critic
version: 2.0.0
verdict: FAIL
coverage: partial (copy only)
for: same-day questionnaires campaign page
outputs:
  page_type: campaign landing page
  critical: 3
  major: 3
  top_fix: make the 3-minute tour the main action
  failed: [M2, M1, C1, F2, F1, P2]
  not_run: [C7, C9, M4, F5, F6, F7]
next: [conversion-research, message-test]
open: [phone screenshot]

What good looks like

  • The first screen is checked before anything below it
  • Every fail quotes the page, and every fix is a concrete change
  • Checks it couldn't run are named, never passed
  • Rewrites sound like you and add no figure you can't source
  • The top 3 fixes come in order, with the reason each comes first

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Does it need screenshots?

No, but they help. From pasted copy it runs the message, proof and friction checks, and marks hierarchy, phone layout and speed as not run.

Why does the page type change the result?

Each page has a different job. A paid campaign page loses money when the promise breaks, while a comparison page loses skeptics when the proof is thin. The weights follow the job.

Can it review a competitor's page?

Yes. It runs every check except the ones that need their sources, then compares their first screen with yours.

Will it tell us to add sections?

Only when a failed check needs one, such as an objection with no answer. Page length alone isn't a finding.