Skip to content

Checks

Message pre-test

Scores a page or ad on clarity, relevance, value and difference for each of your buyer roles, before you pay for a real buyer panel.

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 message-test.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
Message testing
Needs
The page, ad or email, or 2 versions of it · Who reads it, as a role
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedCreative
SkillStrategy and messaging

Message pre-test

Example output

Score a page for one buyer role
CriterionScoreGap
Clarity5None
Value3No measured result
Difference2Fails the swap test

Scores a page or ad on clarity, relevance, value and difference for each of your buyer roles, before you pay for a real buyer panel.

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
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

message-test/SKILL.md

In the folder

  • message-test/SKILL.md
  • message-test/references/rubric.md
  • message-test/scripts/pretest.py
---name: message-testdescription: >-  Pre-tests a B2B SaaS landing page, ad, email or homepage on 4 criteria (clarity, relevance, value and difference)  as each buying group role in the company profile would read it, with every score tied to a quoted line, a  competitor swap test for difference, and fixes drawn from the messaging house. Returns PASS, REVISE, FAIL or  BLOCKED. It is a cheap screen before paying for a real buyer panel, and it says plainly that it does not replace  one: no simulated buyer counts or percentages. Use it before a page or campaign goes live, or to compare 2  versions. Not for tone and banned words (use voice-check), facts and numbers (use claim-check) or a full  conversion review of a page (use landing-page-critic).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads a page if the agent can browse,  otherwise pasted copy or a screenshot's text. Python 3 is optional, for the mechanical signals and the verdict.metadata:  author: SHAKEUP  version: "1.0.0"  stage: check  plugin: strategy-messaging  homepage: https://www.shakeuphq.com/skills/message-test--- # Message pre-test Reads a page or an ad the way each buyer role would, scores what they'd get from it, and says what to fix before real buyers see it. ## Purpose - One job: score one asset (or 2 versions of it) on clarity, relevance, value and difference for the profile's buyers, with fixes.- Stage: check. It ends in a verdict with the rule that fired.- It is a model reading as a buyer, not buyers. It catches clear problems cheaply. It can't say what real buyers think, so it never reports a share of buyers, a preference count or a simulated panel result.- It suggests line fixes from the messaging house. It doesn't rewrite the asset. ## Use when / not | Situation | Use || --- | --- || A page, ad, email or homepage is drafted and you want to know if buyers will get it | this skill || Choosing between 2 versions before a live test or a panel | this skill, both versions, same roles || The messaging house is new and you want to see if a page carries it | this skill || Tone, dashes, banned words | `voice-check` || Every number and claim backed | `claim-check` || Structure, proof, friction and the next step on a page | `landing-page-critic` || The core message itself isn't set | `messaging-house` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The asset: page copy, ad, email (or 2 versions) | yes | the URL if the agent can browse; pasted copy; a screenshot's text | `BLOCKED` || Where the buyer meets it: search ad click, cold feed, retargeting, email, direct | no | the reader's note; inferred from the asset type | infer and state the assumption || Who reads it: the primary role | yes | the reader's note; the champion in `### Buying group` (say you used it) | ask once only when there is neither; `BLOCKED` if still unknown || The buying moment it targets | no | the reader's note; `## Buying moments` | score relevance against the role only; coverage `partial` || Positioning and messaging house | no | `## Positioning` and `## Messaging house` in the profile | difference is scored without a main alternative or marked `not scored`; fixes can't point to pillars | ## Profile read From `company-profile.md`: `## Company` (category), `## Buyers` with `### Buying group` (what each role cares about and their objection), `## Buying moments`, `## Positioning` (alternatives, unique attributes, best fit, category), `## Messaging house`, `## Competitors`, `## Claims and proof`, `### Words to avoid`, and every `## Learnings` entry for `message-test` with status `noted`, `fact` or `accepted`. No profile: ask once for who reads the asset (role and company type) and what they use today instead. Continue with coverage `partial`; say `company-profile`, `positioning` and `messaging-house` make this sharper. ## Steps 1. **Apply learnings.** Accepted learnings for this skill override the defaults (for example a tightened anchor after a panel disagreed). Name each one that fires. 2. **Frame the test.** Write down before reading the asset: asset type, placement, the primary role and up to 2 secondary roles, the moment, and the alternative this reader uses today. Each comes from the reader, the profile, or an assumption labelled `assumed`. 3. **Mechanical pass.** Run:    ```   python3 scripts/pretest.py --asset asset.txt --profile company-profile.md   ```    Mark the fold with a line `---fold---` if you know it; otherwise the first screen is the first 50 words (about a headline, a subhead and a button). Without Python, note by hand: the first screen's text; the category's key words in it; each role named word for word; each moment and alternative with 3 in 4 of its key words present (fewer usually means the category, not the moment); the share of the core message's and each pillar claim's key words present; empty words and words to avoid; every figure; counts of we/our/us against you/your; sentences over 30 words. Say which way you ran it. These are signals for the scores, not scores. 4. **First-screen read.** As the primary role, from the placement, read only the first screen (the whole ad for an ad). Write one line each, using only words on the asset:   - What is it?   - Who is it for?   - What do I get?   - What do I do next?   - What proves it, and where does the proof sit (first screen, below it, or nowhere)?    A line you can't write from the text is a clarity gap; proof below the first screen is a value note. Don't fill it from the profile: the buyer hasn't read the profile. 5. **Full read.** Read the whole asset as the role, at the moment. List the questions this role would still have, starting from their objection in the buying group and the worries in the buying moments. A question the asset answers is crossed off with the line that answers it. 6. **Run the 2 tests.**   - **Swap test (difference):** put each top alternative's name in place of the company's name in the headline and the 3 strongest lines. A line still true of the alternative doesn't count as difference. The status quo counts as an alternative ("a spreadsheet does this too").   - **Measure test (value):** is the promised result something this role measures (from "cares about" in the buying group)? Is it big enough to beat the comfort of the current way? 7. **Score each criterion 1 to 5** against the anchors in `references/rubric.md`, per role. Every score cites the line or lines it rests on and the gap that held it back.    | Criterion | The question |   | --- | --- |   | Clarity | Can the buyer say what it is, who it's for, what they get and what to do, from the first screen? |   | Relevance | Does it name this role's situation or moment, in words they use? |   | Value | Is the result one they measure, with proof, and does it answer their main worry? |   | Difference | After the swap test, is there a reason to pick this over what they use today? |    Write `not scored` when the input it needs is missing (difference with no known alternative and none the reader can name). Never guess a score to fill the gap. 8. **Get the verdict.** Run `python3 scripts/pretest.py --scores scores.json`, or apply the table below by hand. The primary role decides it; secondary roles' gaps are notes. 9. **Write the fixes.** For every criterion at 3 or below: the line to change, what is missing, and where to draw the fix from (a pillar, the core message, a unique attribute, a buying moment phrase). Draw only on pillar parts that have proof; leave out a pillar's comparison or figure that has no source. Never add a claim missing from `## Claims and proof`; send every figure on the asset to `claim-check`. 10. **Write the panel brief.** What a real buyer test should ask: the audience (role, seniority, company type from the ideal customer, or from what the reader said), the versions, the 4 questions in step 7 in the buyer's terms, and the decision each answer drives. Options: a paid B2B message-testing panel, or 5 interviews with buyers who match the ideal customer. 11. **Comparing 2 versions:** score both for the same roles. Report the criteria where they differ by 2 or more points. A difference of 1 is within what a re-read changes, so call it a tie and let the panel decide. ## Check and verdict | Verdict | Condition, for the primary role (first match wins) || --- | --- || `BLOCKED` | No asset; or no buyer role after asking once || `FAIL` | Any criterion scores 1 or 2 || `REVISE` | Any criterion scores 3, or any criterion is `not scored` || `PASS` | Every criterion scores 4 or 5: ready for a real buyer panel or a live test, not proof that buyers agree | When clarity is 2 or less, mark the other scores provisional and put the clarity fix first: a buyer who can't tell what it is never reaches the value. Named stop: `BLOCKED` asks for the asset or the role. No other stops; the skill that made the asset decides what to do with a `FAIL`. Every output carries this line, unchanged in meaning: "This is a pre-test by a model reading as your buyers, not a test with buyers. Use it to fix clear problems before you pay for a real panel; don't report it as buyer research." ## Learn Append to `## Learnings`: - a real panel or a live test comes back for an asset this skill scored: a `fact` with both results side by side and the source;- the panel scores a criterion 2 or more points below this skill on 2 assets: a `proposal` to tighten that criterion's anchor for this company;- the reader rejects a finding: a `proposal` with their reason. Example: `- 2026-10-24 | message-test | proposal | Score clarity 1 point lower when the first screen names no card provider | Why: panel clarity 2.1 vs pre-test 4 on 2 pages (buyer panel, 2026-10) | Status: proposed` ## Output ```Verdict: <PASS | REVISE | FAIL | BLOCKED> · Coverage: <complete | partial (what's missing) | blocked> · Signals: <script | by hand>This is a pre-test by a model reading as your buyers, not a test with buyers. ... Frame: <asset type> · <placement> · primary role <role> (<from profile | reader | assumed>) · moment <moment> · alternative <alternative> First screen, as <role>What is it: <line or "can't tell">Who is it for: <line or "can't tell">What do I get: <line or "can't tell">What do I do next: <line or "can't tell">What proves it: <proof, and where it sits, or "none"> | Criterion | <primary role> | <secondary role> | Line it rests on | Gap || Clarity | n | n | "<quote>" | ... || Relevance | ... || Value | ... || Difference | ... (swap test: <lines that survived>) | Questions the buyer still has: <list, with the objection each comes from> Fixes (lowest criterion first)| Line | Criterion | What's missing | Draw from | Figures for claim-check: <list>Panel brief: <audience, versions, 4 questions, decision each drives>Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: message-testversion: 1.0.0verdict: REVISEcoverage: partial (no buying moment named; placement assumed)for: <asset name>outputs:  primary_role: Head of finance  scores: {clarity: 5, relevance: 4, value: 3, difference: 4}  lowest: value  fixes_ready: true  figures_to_check: ["3 days"]next: [claim-check, landing-page-critic]open: [run a buyer panel before scaling spend]``` Made-up values for illustration. The skill that made the asset reads `lowest` and the fix table, rewrites, and runs this check again (its own loop, at most 2 rounds). ## Failure modes | Problem | Do this || --- | --- || The reader asks for "what % of buyers would" or a simulated panel | Refuse that part plainly; give the scores and the panel brief instead || The asset is a screenshot with no text available | Ask for the copy; don't score from a description of an image || The role isn't in the profile's buying group | Score for the stated role, label it `assumed`, and suggest `buying-moments` || A score depends on knowledge only the company has | The buyer doesn't have it: score from the asset alone || The page is long | Score the first screen for clarity; the full read for the rest; say how much you read || 2 versions score the same | Call it a tie; the panel or the live test decides | ## Influenced by - Wynter's message testing (Peep Laja), which asks real B2B buyers from a verified panel to rate a page's clarity, relevance, value and differentiation: https://wynter.com/post/message-testing- What we changed: the same 4 questions are answered by a model reading as each role in the company profile, with anchored scores tied to quoted lines, a competitor swap test, a first-screen read and a verdict rule. It is a screen before a real panel, never a substitute, and it never reports buyer counts.

How it thinks

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

What we built

We kept the 4 questions buyers answer in a real message test, and added anchored scores tied to quoted lines, a first-screen read and a competitor swap test. It's a screen before a panel, and it never reports buyer counts or percentages.

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

    Read your learnings first

    Panel results you recorded against past pre-tests can make its scoring stricter for your company.

  2. 02

    Frame the test

    Asset, placement, the role reading it, the moment and what that buyer uses today. Each comes from you, your profile or a stated assumption.

  3. 03

    Read the first screen as the buyer

    It answers what it is, who it's for, what they get and what to do next, using only words on the page. The buyer hasn't read your profile.

  4. 04

    Swap in a competitor

    Each strong line is reread with an alternative's name in place of yours. A line that stays true doesn't count as difference.

  5. 05

    Score 4 criteria from 1 to 5

    Clarity, relevance, value and difference, each against written anchors and tied to the line it rests on. Each role gets its own scores.

  6. 06

    Fix from the messaging house

    Every weak line gets a fix drawn from your pillars or positioning, never a new claim. Figures go to the claim check.

  7. 07

    Brief the real panel

    It writes the audience, the versions and the questions for a buyer panel or 5 buyer interviews, and says what each answer decides.

What you need

  • Required

    The page, ad or email, or 2 versions of it

    The page address, if your agent can browse → Copy you paste → The text from a screenshot

  • Required

    Who reads it, as a role

    Your note → The buying group in your company profile, with the champion by default

  • Optional

    Where the buyer meets it, such as a search ad click or a cold feed

    Your note → Inferred from the asset

  • Optional

    Your buying moments, positioning and messaging house

    The matching sections of your company profile

What it can return

  • PASS

    Every criterion scores 4 or 5 for the main reader. It's ready for a real panel or a live test, not proof that buyers agree.

  • REVISE

    A criterion scores 3, or couldn't be scored for lack of input.

  • FAIL

    A criterion scores 1 or 2 for the main reader.

  • BLOCKED

    There's no asset, or no reader role after one question.

Where it stops for you

  • When there's no asset or no reader role after one question

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 expense software. The page is where its search ads land, and the main reader is a head of finance.

Verdict

REVISE

Coverage: complete

This is a pre-test by a model reading as your buyers, not a test with buyers. Use it to fix clear problems before you pay for a real panel, and don't report it as buyer research.

CriterionHead of financeIT reviewerLine it rests onGap
Clarity54"Close the books in 3 days"None for finance
Relevance42"matches every receipt to the card feed"No audit moment named
Value33"Close the books in 3 days"No proof on the page
Difference43"matches every receipt to the card feed"Survives the swap test
First screen, as the head of finance

it's expense software for finance teams, and the result is books closed in 3 days. The next step is a demo.

Fixes

put the source for "3 days" beside the headline, or cut the figure until the claim check finds one. Add the audit moment in the subhead, from the buyer's own words in the messaging house.

Panel brief

15 heads of finance at companies with 200 to 1,000 employees. Show the page and ask what it is, who it's for, what they'd get, and why they'd pick it over their spreadsheet.

The handoff block the next skill reads
skill: message-test
version: 1.0.0
verdict: REVISE
coverage: complete
for: expense software search landing page
outputs:
  primary_role: Head of finance
  scores: {clarity: 5, relevance: 4, value: 3, difference: 4}
  lowest: value
  figures_to_check: ["3 days"]
next: [claim-check, landing-page-critic]
open: [run the panel before scaling search spend]

What good looks like

  • First-screen answers written only from words on the page
  • A quoted line behind every score
  • Difference checked with a competitor's name swapped in
  • Fixes that come from your messaging house, with no new claims
  • The plain line that this isn't buyer research

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 this replace a buyer panel?

No. A model reading as your buyers catches clear problems cheaply, like a headline nobody could repeat. Only real buyers can tell you what they think, so it writes the panel brief too.

Why score each role on its own?

A champion and a blocker read the same page differently. The main reader decides the verdict, and the other roles' gaps become notes.

What's the swap test?

Each strong line is reread with a competitor's or the status quo's name in place of yours. If the line is still true, it doesn't set you apart.

Can it compare 2 versions?

Yes. It scores both for the same roles and only calls a winner on a gap of 2 points or more. Smaller gaps go to the panel.