The stranger test

What it is and why it matters

One of our posts went past a hundred thousand views. Underneath, a comment: “did anyone understand anything?”. And all those views brought no downloads.

The problem wasn't the story. It was that we knew too much. You reread your text and it seems clear, because you already have everything else in your head: the product, the why, the bit you cut to keep it short. The reader only has the text.

The stranger test is this: you give only the text to a brand-new agent, with no context and on the smallest model. You ask it three questions: what happened (or what this product does), who it's for and why it concerns me, what I should do now. Then you ask for the words and passages it didn't understand. If it gets an answer wrong, the problem isn't the agent: you rewrite. It works for anything someone else will read: posts, landing pages, App Store pages, onboarding, emails.

1. The three-question prompt

In place of [paste here], put the text, only the text: no file name, no explanations of your own.

You're an ordinary person scrolling on your phone. You know nothing about who wrote this text, or about their product. Read only the text below and answer:

1. What happened? Or, if it's a product: what does it do?
2. Who is it for? Why should I care?
3. What should I do now?

One or two sentences per question, in your own words. If the answer isn't in the text, write “I don't know”. Don't guess. If you had to work something out, say so.

Then list:
- the words you didn't understand;
- the passages where you stopped or had to reread, with the exact sentence.

Don't give advice on how to improve it.

TEXT:
[paste here]

The first question changes a little with the kind of text:

• a post that tells a story: what happened, and what the product has to do with it;
• a landing page or an App Store page: what this product does;
• an onboarding screen: what I'm supposed to do on this screen;
• an email: what they're asking me for.

“Don't give advice” isn't a detail. If you leave it free, it rewrites your text, and you no longer know where it got lost.

2. Run it as a subagent, on the small model

In Claude Code a subagent starts with a fresh context: it doesn't see your conversation. But careful: by default it reads your CLAUDE.md files, which is exactly where your product is described. At that point it's no longer a stranger. That's why I create a permanent agent, in the file .claude/agents/stranger.md:

---
name: stranger
description: Reader with no context. Reads only the text it's given and says what it understood. Use it before publishing any text.
tools: Read
model: haiku
omitClaudeMd: true
---

You're an ordinary person scrolling on your phone. You know nothing about who wrote the text, or about their product. Read only the text or the file you're given, nothing else.

Answer:
1. What happened? Or: what does this product do?
2. Who is it for? Why should I care?
3. What should I do now?
One or two sentences per question, in your own words. If the answer isn't in the text, write “I don't know”. If you had to work it out, say so.

Then list the words you didn't understand and the passages where you stopped, with the exact sentence. Don't give advice.

• model: haiku is the smallest model. If it understands, so does someone scrolling in a hurry.
• omitClaudeMd: true starts it without your CLAUDE.md files. It needs a recent version of Claude Code.
• tools: Read is enough for it to open a file or a post's images. No web, no edits.

Then, in the session, write this (or type @ and pick it from the list):

Use the stranger subagent, with the haiku model, on the text of [file]. Wait for its report and give it to me as it is, without commenting on it.

Why do I repeat “with the haiku model”? When Claude launches a subagent it can pick a model for it, and that choice beats the file's model field. Writing it in the message removes the doubt. If you don't want to create the file, you can ask for a general-purpose subagent with the haiku model: it works, but it reads your CLAUDE.md files, so the test is worth less.

3. Turn it into a command

If you use it often, a skill turns it into a command: /stranger followed by the file. Create .claude/skills/stranger/SKILL.md:

---
name: stranger
description: Stranger test on a text before publishing it.
disable-model-invocation: true
context: fork
agent: stranger
background: false
---

Read only $ARGUMENTS and run the test: the three questions, then the words and passages you didn't understand.

• context: fork runs it in a separate subagent; agent: stranger picks your agent, with its model and without CLAUDE.md.
• background: false makes it wait for the report, instead of sending it later as a notification.
• disable-model-invocation: true means it only starts when you call it.

You use it like this: /stranger landing/index.html. The report comes back into your conversation.

4. How to read the answers

Before I run it, I write the three right answers myself, one line each. Then I compare:

• Right answer, in its own words: it passes.
• Wrong answer or “I don't know”: that piece isn't in the text, or it can't be understood. Rewrite.
• Right answer but worked out: for me it's wrong. Someone scrolling doesn't stop to work things out.
• Words it didn't understand: each one gets swapped for an everyday word, or explained in the same sentence.
• Passages where it stopped: usually a link is missing between two sentences, or two sentences contradict each other.

A real example, from one of our posts. On the first round the agent got stuck on words like “physiologist” and “lactate”. By the fourth it answered correctly, but about the link to the app it wrote that it had worked it out itself: we rewrote that too. By the sixth it answered correctly without working anything out.

5. What to fix, and when to run it again

• Fix the text, not the prompt. If you add explanations to the prompt, the real reader won't have them.
• Every technical term gets translated in the same sentence: not “VO2max”, but “your breath”.
• A number on its own says nothing: put its comparison next to it. And never two percentages to compare with each other.
• Every round with a new agent. The previous one has already read the old version: it's no longer a stranger.
• You stop when the three answers are right without working anything out and no stumbles remain on the passages that matter.

For us the record goes into the day's report, and a post without one doesn't even reach the final critic (the whole system is at strikt.fit/marketing-notturno).

Common mistakes

• Asking Claude “is it clear?” in the same chat: it knows everything you know, and it'll say yes.
• Letting it read your CLAUDE.md files: it already knows the product. Use omitClaudeMd: true.
• Giving it the text wrapped in your own explanations: give it only what the reader will see.
• Using a big model: it fills in even what's missing from the text.
• Arguing with it: if it didn't understand, you didn't write clearly. Rewrite.

Checklist

• The three right answers written by me, before the test
• Agent with model: haiku and omitClaudeMd: true
• I gave it only the text, with no context
• Three right answers, without working anything out
• Every hard word replaced or explained in the same sentence
• Last round done with a new agent

Frequently asked questions