Lesson 4 of 32 · Part One, the principles

The Law of the Master Persona

Telling the model who to be changes what it notices. A village health trainer, a senior auto driver and a YouTube script editor would answer the same question in different words and also disagree about what the problem is. The useful part of a persona is not the voice - it is the priorities, the examples and the cautions that come with doing that job for years.

Reach for it whenWhen the answer has to sound like someone your listeners already trust - a trainer, an elder, an editor - and has to carry that person's practical warnings as well as their advice.

How to use The Law of the Master Persona in Claude

Give the persona its own tag with four things inside: the role, what it cares about, how it speaks, and what it will not do. That last line is what makes a persona trustworthy rather than theatrical - the health educator who sends clinical questions to the ANM, the mentor who will not quote an income he cannot stand behind. Then ask for the verdict as well as the content: Claude will tell Neha which hook it would refuse to publish if you ask it to, and that refusal is the expert part.

  • One tag: role, priorities, voice, and what it will not do.
  • Ask what the expert would refuse, not only what it advises.
  • Say "be direct, do not soften this" - it will comply.

From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.

Give the persona its own tag with four things inside: the role, what it cares about, how it speaks, and what it will not do. That last line is what makes a persona trustworthy rather than theatrical - the health educator who sends clinical questions to the ANM, the mentor who will not quote an income he cannot stand behind. Then ask for the verdict as well as the content: Claude will tell Neha which hook it would refuse to publish if you ask it to, and that refusal is the expert part.

From the book Write your persona into a Project's instructions so every chat in it starts with that expert. Use Styles to choose or create a voice, such as warm and simple, that suits your audience, and edit an earlier message to test a sharper persona.
Worth knowingClaude leans towards being agreeable, so a persona can come out encouraging where you needed it blunt. Say inside the tag that a polite answer is no use to you, and ask explicitly for what the expert would refuse, object to or send elsewhere.
  • One tag: role, priorities, voice, and what it will not do.
  • Ask what the expert would refuse, not only what it advises.
  • Say "be direct, do not soften this" - it will comply.
  • Put a standing persona in a Project's instructions.
  • Use Styles to hold one voice across a whole Project.
A Plain Answer vs. A Trusted Health Educator

Instead of

Tell me about breastfeeding for my village talk.

Claude version

<persona>
You are a community health educator who has trained ASHA workers in rural Rajasthan for twenty years.
What you care about: that the mother remembers three things tomorrow morning, and that nobody in the family is shamed in front of the others.
How you speak: simple spoken Hindi, short sentences, examples from village life, no medical vocabulary.
What you never do: answer a medical question yourself. You send it to the ANM or the doctor, warmly, not as a disclaimer.
</persona>

<my_audience>
New mothers in a village near Bhilwara, Rajasthan. Mothers-in-law and sisters-in-law will be sitting there too, and will have opinions. Many of the women have not studied beyond school. The talk is ten minutes, outdoors, with noise.
</my_audience>

<must_cover>
Why breastfeeding in the first six months matters.
Three myths common in these villages, each with a gentle way to answer it.
The reminder to take any problem to the ANM or the doctor.
</must_cover>

<if_unsure>
Where something is a matter of clinical advice rather than general health education, say "this is for the ANM to answer" instead of answering it. I would rather have a gap in my talk than tell a mother something wrong.
</if_unsure>

<task>
Write the talk in simple Hindi, as it would be spoken.
Before the talk, two lines in your own voice as the educator: which part of this talk most often goes wrong in front of a family, and what you do about it.
</task>
Open Claude 1,438 characters

Then loop it

  1. That is the right voice. Keep the three points and the myths. Add a short story about a worried mother-in-law who believes her daughter-in-law's milk is not enough - bring her along, do not correct her. Tell me if the story pushes the talk past ten minutes.
  2. Now five picture-card captions in Hindi, one line each, in an Artifact so I can edit them before printing. Then put this persona into a Project called "Village health talks" so every topic starts with the same educator.

Why it worksThe educator persona sets the tone, vocabulary and caution level, so the talk sounds like a trainer, not a textbook.

A Generic Reply vs. An Experienced Mentor

Instead of

How can auto drivers earn more?

Claude version

<persona>
You are a senior auto driver in Chennai and a union mentor, twenty years of guiding younger drivers.
What you know that a general adviser does not: what the apps actually pay after commission and incentives, what waiting in the airport queue costs in hours, which school routes are worth having and which parents are trouble.
How you speak: like an elder brother. Direct. You do not oversell. If an idea is wrong for a particular driver, you say so.
What you will not do: quote a monthly figure you cannot stand behind. Where it depends on the man and the route, you say that instead.
</persona>

<about_me>
I own my auto. I earn mainly from local stand rides in Chennai. I read Tamil more easily than English.
</about_me>

<the_three_options>
1. Joining ride apps.
2. Airport and tourist trips.
3. Regular school trips.
</the_three_options>

<format>
For each option: the good side, the hard side, and the first step I can take next week.
Then your own verdict - which one you would start with in my position, and which you would tell me to leave for now.
</format>

<if_unsure>
Where earnings depend on the route, the season or the app's current commission, say so rather than giving me a number I will plan around. Mark anything that changes often as "ask at the stand".
</if_unsure>

<task>
Advise me, in simple Tamil, as the mentor you are.
Before the three options, one line on what you noticed about my situation that I did not tell you directly.
</task>
Open Claude 1,471 characters

Then loop it

  1. Keep the school trips section. Now, still as the mentor, tell me what to say to parents to earn their trust - what they are actually afraid of when a stranger drives their child. Tell me honestly if you think I am not ready for school trips yet.
  2. Now a short Tamil message for the parents' WhatsApp group, under 60 words, in an Artifact so I can change it for each school. Then save this mentor's description in a Project called "My auto work".

Why it worksAn experienced-mentor persona brings practical, ground-level advice instead of a generic list.

Flat Ideas vs. A Seasoned Script Editor

Instead of

Give me YouTube video ideas.

Claude version

<persona>
You are a YouTube script editor who has taken Indian food channels from ten thousand subscribers to past one lakh.
What you judge by: whether the viewer is still there at fifteen seconds, and whether they click the next video.
What you are allowed to dismiss: production quality, equipment, thumbnail styling, anything that is not the opening.
How you speak: blunt. You tell a creator when an idea is weak, and you name the one you would refuse to publish.
</persona>

<my_channel>
Quick Maharashtrian recipes for working women. Hinglish, the way I actually speak. About fifteen thousand subscribers, Pune.
Next video: a 10-minute poha recipe.
</my_channel>

<what_i_want>
Three different 15-second opening hooks, one in each style: a curious question, a surprising fact, a relatable problem.
For each: the exact spoken words, and one line on what is on screen while I say them.
</what_i_want>

<rules>
Hinglish, natural. No "hey guys, welcome back to my channel".
Fifteen seconds means fifteen seconds spoken aloud, not read on a page.
</rules>

<task>
Write the three hooks.
Then, as the editor, pick one and say why - and name the one you would refuse to let me publish, and what is wrong with it. Do not soften this; a polite editor is no use to me.
</task>
Open Claude 1,271 characters

Then loop it

  1. Hook two. Keep its shape and rewrite it more casually - me talking to a friend across the kitchen counter, not presenting. Still under fifteen seconds spoken. If making it casual weakens the hook, say so.
  2. Now five thumbnail texts of four words or fewer that match hook two, in an Artifact so I can try them against each other. Then put this editor into a Project called "My channel" so every video starts with the same judgement.

Why it worksThe script-editor persona focuses the AI on hooks, pacing and audience - the things an expert would judge.

Questions

What is The Law of the Master Persona in prompt engineering?

Telling the model who to be changes what it notices. A village health trainer, a senior auto driver and a YouTube script editor would answer the same question in different words and also disagree about what the problem is. The useful part of a persona is not the voice - it is the priorities, the examples and the cautions that come with doing that job for years.

When should I use this technique?

When the answer has to sound like someone your listeners already trust - a trainer, an elder, an editor - and has to carry that person's practical warnings as well as their advice.

How do I apply it in Claude?

Give the persona its own tag with four things inside: the role, what it cares about, how it speaks, and what it will not do. That last line is what makes a persona trustworthy rather than theatrical - the health educator who sends clinical questions to the ANM, the mentor who will not quote an income he cannot stand behind. Then ask for the verdict as well as the content: Claude will tell Neha which hook it would refuse to publish if you ask it to, and that refusal is the expert part.

What are the limits of this technique in Claude?

Claude leans towards being agreeable, so a persona can come out encouraging where you needed it blunt. Say inside the tag that a polite answer is no use to you, and ask explicitly for what the expert would refuse, object to or send elsewhere.

How do I improve the answer after the first try?

After the first answer, check whether the voice fits the role you gave. Does the "mentor" sound practical, or like an encyclopaedia? If not, sharpen the persona: add years of experience, the audience it serves, or how it speaks. You can also switch roles mid-chat: "Now answer as a worried parent - what would you object to?" Once a persona works well, save it so you can call it up again in one click.

Loop it

After the first answer, check whether the voice fits the role you gave. Does the "mentor" sound practical, or like an encyclopaedia? If not, sharpen the persona: add years of experience, the audience it serves, or how it speaks. You can also switch roles mid-chat: "Now answer as a worried parent - what would you object to?" Once a persona works well, save it so you can call it up again in one click.

This lesson on the other tools

The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.