Lesson 25 of 32 · Part One, the principles
The Formation of the Master Mind
This is the chapter that gathers up all the others. Stop treating the machine as a vending machine that takes a request and dispenses an answer, and treat it as a junior partner: it holds more organised knowledge than you ever will and has no aim of its own, while you hold the aim, the facts of your own life, and the judgement about what is actually true and what will actually work. So say what each of you brings, invite it to ask you questions before it drafts anything, and read every draft as the senior partner - this part fits, this part does not, and here is the thing only I could know.
How to use The Formation of the Master Mind in Claude
Write the partnership down. A <my_role> tag and a <your_role> tag at the top of the prompt do more work than any amount of friendly phrasing, because they tell it which decisions are not its to make - and this is the chapter where that matters most. Then put the long material first and the instruction last, as always: the paper, your notes, the facts of your week, and the actual ask at the end. Three mechanics carry a partnership across days rather than one chat. A Project holds the standing instructions and the files - the uploaded study, your constraints, the plan as it stands - so every new chat opens with the shared background instead of your retyping it. An Artifact gives the two of you one living document to correct, which is The Law of the Loop with somewhere to put the work. And when a round goes wrong, editing your earlier message lets you take the partnership down a different path cleanly instead of arguing your way back.
- Write <my_role> and <your_role> so the decisions stay yours.
- Build disagreement into the brief, then ask for it again.
- Keep the standing facts and files in a Project, not in one chat.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Write the partnership down. A <my_role> tag and a <your_role> tag at the top of the prompt do more work than any amount of friendly phrasing, because they tell it which decisions are not its to make - and this is the chapter where that matters most. Then put the long material first and the instruction last, as always: the paper, your notes, the facts of your week, and the actual ask at the end. Three mechanics carry a partnership across days rather than one chat. A Project holds the standing instructions and the files - the uploaded study, your constraints, the plan as it stands - so every new chat opens with the shared background instead of your retyping it. An Artifact gives the two of you one living document to correct, which is The Law of the Loop with somewhere to put the work. And when a round goes wrong, editing your earlier message lets you take the partnership down a different path cleanly instead of arguing your way back.
From the book Make a Project for a long partnership, such as a business plan or a family routine, and add standing instructions and files like your notes or the uploaded paper. Every chat in that Project starts with your shared background.
- Write <my_role> and <your_role> so the decisions stay yours.
- Build disagreement into the brief, then ask for it again.
- Keep the standing facts and files in a Project, not in one chat.
- Demand the section or table behind every number from a document.
- Prune the Project: abandoned constraints keep steering answers.
Instead of
Make a daily schedule for me and my son.
Claude version
<my_role> I bring the facts of my life and the final judgement on what will survive a real week. I am the one who has to live inside this routine. </my_role> <your_role> You bring structure, and options I have not thought of. You do not decide what matters to us. </your_role> <fixed> Office 9:30 to 6. One hour's commute each way. My son is 8. School bus leaves 7:40, returns 3:30. A neighbour keeps him from 3:30 until I get home. </fixed> <my_aim> A weekday routine that leaves the two of us 30 real minutes together every evening, and does not depend on me having more energy than I actually have. </my_aim> <how_we_work> First, ask me the five questions whose answers would most change the plan. Nothing else in that message - no draft, no suggestions. Then one draft routine as an Artifact, which we will correct together rather than replace. With the draft, tell me which two parts you are least confident about, and name anything in what I have told you that strikes you as unrealistic. I would rather be contradicted now. </how_we_work>
Then loop it
- Answers: he loves cricket, homework takes about 40 minutes, and I cannot cook every night. Now draft the routine.
- Tuesday will not work - I have late calls. Adjust Tuesday, and suggest three 20-minute dinners I can cook in bulk twice a week.
Why it worksDeepa supplies the real constraints and judgement; the AI supplies structure and ideas, and they shape the plan together.
Instead of
Tell me how to sell furniture online.
Claude version
<me> Carpenter in Saharanpur, twenty years of hand-carving in sheesham. I carve all day; I have no shop assistant and no English typing speed. </me> <my_vision> To sell small carved pieces - jewellery boxes, wall panels, pooja stands - directly to customers in the big cities, without middlemen. This is settled. It is not a proposal for you to improve. </my_vision> <my_role> Wood, carving, timber, finish, how long a piece takes, what quality I will not compromise. The final choice is mine. </my_role> <your_role> Online selling, which I know nothing about. Be the partner who knows that side. Do not explain carpentry to me. </your_role> <this_message> Three genuinely different ways to begin - different in approach, not three versions of one. For each: hours a week it needs from me, rupees needed up front, what can go wrong, and how long before a first sale is realistic. Then stop and ask me which fits my life. We will plan that one together, step by step, the way a chain of thought is built - one link approved before the next. </this_message> <be_honest> If one of your three is plainly wrong for a man who is at the bench all day, say so instead of listing it to make up the number. </be_honest>
Then loop it
- Option 2, selling on Instagram with my nephew's help, feels right. Ask me what you need to know to plan the first month.
- Here are my answers. Make a 4-week plan, and keep Sundays free - that is my family day.
Why it worksShakeel keeps the vision and the final choice; the AI contributes the knowledge he lacks.
Instead of
Make my journal club presentation.
Claude version
<the_paper> [Upload the paper here, or add it to this Project. It contains no patient data. The long material goes first; my instruction comes last.] </the_paper> <me> Junior resident in Chennai. Fifteen-minute journal club next Thursday, in front of my department. </me> <my_role> I decide the key message and I judge whether the findings apply to the patients we actually see. That judgement is mine and I am not delegating it. </my_role> <your_role> Structure, clarity, and a careful second reading of the paper. </your_role> <step_one> Summarise in plain numbered points: study design, sample size, follow-up period, primary outcome, main results. Against each point, name the section, table or figure you took it from, so I can check every number against the paper myself. Stop there and wait for me. </step_one> <step_two> After I confirm the figures: a ten-slide outline, one line per slide, as an Artifact we can keep editing. </step_two> <flag> Mark anything you are not certain the paper states - including the things that look obvious. And when I give you my key message, do not simply arrange the slides around it: tell me what in the results supports it and what sits against it. </flag>
Then loop it
- Your summary has the follow-up period wrong - it is 12 months, not 6. Fix that, then continue.
- My key message: these results may not apply to the patients our hospital sees. Rework slides 8 to 10 around that, and give me three questions my professor may ask.
Why it worksClear roles - she judges, the AI organises - keep the expert in charge and the AI checkable.
Questions
What is The Formation of the Master Mind in prompt engineering?
This is the chapter that gathers up all the others. Stop treating the machine as a vending machine that takes a request and dispenses an answer, and treat it as a junior partner: it holds more organised knowledge than you ever will and has no aim of its own, while you hold the aim, the facts of your own life, and the judgement about what is actually true and what will actually work. So say what each of you brings, invite it to ask you questions before it drafts anything, and read every draft as the senior partner - this part fits, this part does not, and here is the thing only I could know.
When should I use this technique?
On the work that matters and will take more than one sitting - a routine for your family, a plan to sell what you make, a presentation your professor will question - where the answer has to carry your judgement and not just the average of everyone else's.
How do I apply it in Claude?
Write the partnership down. A <my_role> tag and a <your_role> tag at the top of the prompt do more work than any amount of friendly phrasing, because they tell it which decisions are not its to make - and this is the chapter where that matters most. Then put the long material first and the instruction last, as always: the paper, your notes, the facts of your week, and the actual ask at the end. Three mechanics carry a partnership across days rather than one chat. A Project holds the standing instructions and the files - the uploaded study, your constraints, the plan as it stands - so every new chat opens with the shared background instead of your retyping it. An Artifact gives the two of you one living document to correct, which is The Law of the Loop with somewhere to put the work. And when a round goes wrong, editing your earlier message lets you take the partnership down a different path cleanly instead of arguing your way back.
What are the limits of this technique in Claude?
Claude is an unusually agreeable junior partner, and in this chapter that is the principal risk rather than a footnote. Announce your key message and it will organise the evidence around it; declare that your routine will hold and it will tell you why it should. So write disagreement into the brief - name where you think I am wrong, and say what would have to be true for this to fail - and ask it again after the draft. Two harder limits. It cannot judge whether a study's findings apply to the patients in your hospital; that judgement is the senior partner's and cannot be delegated, however fluent the answer sounds. And it has no way of knowing whether what it has told you about a document is what the document says - so when the paper is uploaded, insist on the section or table behind each number and check them yourself. A long Project also accumulates: a constraint you abandoned in February can still be shaping answers in June, so prune the files.
How do I improve the answer after the first try?
In a Master Mind, the loop is the partnership itself. Do not expect a finished product in one message. Share your aim and invite the AI to ask questions first. Read its draft as the senior partner: what fits your vision, what is wrong, what only you would know? Say so plainly - "This part works, this does not, here is why" - and let it rebuild. Each round adds your judgement to its knowledge. When a plan is settled, save it and the key facts so the next session starts where you left off.
Loop it
In a Master Mind, the loop is the partnership itself. Do not expect a finished product in one message. Share your aim and invite the AI to ask questions first. Read its draft as the senior partner: what fits your vision, what is wrong, what only you would know? Say so plainly - "This part works, this does not, here is why" - and let it rebuild. Each round adds your judgement to its knowledge. When a plan is settled, save it and the key facts so the next session starts where you left off.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.