Lesson 16 of 32 · Part One, the principles
The Seed of a New Idea
Brainstorming
Do not ask for the perfect name. Ask for thirty rough ones. Volume and range are what the machine is genuinely better at than you - it will lay out more options in a minute than you would reach in a week - while the judgement stays where it belongs, with the person who knows the shop, the buyers and the budget. State a number, demand variety in named directions, forbid it from choosing a winner, and then do the choosing yourself.
How to use The Seed of a New Idea in Claude
Put the quotas in a tag and they hold: so many in this direction, so many in that, so many that are deliberately risky. Ask for the output as a table in an Artifact and you get something you can actually work in - delete rows, keep rows, download the shortlist - instead of a list buried four messages back. One extra column does more than any other instruction: have it mark each idea obvious, decent or strange, by its own judgement. You will disagree with some of those marks, and the disagreement tells you where your own taste sits. When a whole direction is wrong, edit your first message rather than arguing downward from a bad list.
- Quotas in a tag - so many per direction, so many deliberately risky.
- Table in an Artifact: keep rows, cut rows, download the shortlist.
- Add a column where it marks each idea obvious, decent or strange.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Put the quotas in a tag and they hold: so many in this direction, so many in that, so many that are deliberately risky. Ask for the output as a table in an Artifact and you get something you can actually work in - delete rows, keep rows, download the shortlist - instead of a list buried four messages back. One extra column does more than any other instruction: have it mark each idea obvious, decent or strange, by its own judgement. You will disagree with some of those marks, and the disagreement tells you where your own taste sits. When a whole direction is wrong, edit your first message rather than arguing downward from a bad list.
From the book Ask for the ideas as a table in an Artifact, so you can see them side by side, edit them and download your shortlist. To try a different direction, edit your first message and run it again.
- Quotas in a tag - so many per direction, so many deliberately risky.
- Table in an Artifact: keep rows, cut rows, download the shortlist.
- Add a column where it marks each idea obvious, decent or strange.
- Ask what it ruled out as too silly to include.
- Wrong direction? Edit the first message instead of arguing downward.
Instead of
Give me the perfect name for our saree collection.
Claude version
<the_cooperative> A handloom weavers' cooperative in Pochampally, Telangana. Our new collection is ikat cotton sarees, sold online. The buyers we are after are working women of 25 to 45 in metros who like handmade, natural things and understand what they are paying for. </the_cooperative> <quotas> Thirty names in all. Ten in plain English that a Mumbai buyer grasps instantly. Ten built on a Telugu word, with the word and its meaning given. Ten centred on the weaver, the loom or the hands that made it. </quotas> <forbidden> Silk, in any name - this is cotton. Ethnic, Fusion, Heritage, Revival. Anything that would sit comfortably on a mass-market label. </forbidden> <judgement_stays_with_me> Do not choose. Do not rank. Do not tell me which you prefer, and do not order them so that your preference shows. I will decide, and I want a clean eye when I do. </judgement_stays_with_me> <task> Build an Artifact: a table with columns for number, name, direction, tagline under eight words, and a last column where you mark each one obvious, decent or strange - your honest assessment, not a flattering one. Then, in your reply and not in the Artifact, list the five names you drafted and discarded as too odd to include. I want to see them. </task>
Then loop it
- Keep 4, 12 and 23 in the Artifact and cut the rest. Then add ten more in their spirit - and tell me what those three share, since your reasoning may catch something I have only felt.
- Two of your discarded names were better than most of the thirty. Build ten more in that direction, and stop protecting me from strange ones.
- Combine the Telugu quality of 12 with the brevity of 4. Five options, with pronunciation for a buyer who does not read Telugu, and mark which of the five you think a weaver would be happy to have on his work.
Why it worksIt asks for 30 seeds in three directions and keeps the choosing - the human judgement - with Padma.
Instead of
Suggest a theme for a 60th birthday party.
Claude version
<the_brief> Event planner in Jaipur. A 60th birthday party for a retired railway officer who loves old Hindi film songs and train journeys. His family are my clients; they are fond and slightly cautious. </the_brief> <hard_constraints> Budget ₹2.5 lakh, everything included. One hundred and twenty guests, from 5 years old to 85. A garden lawn, February evening - cold after 8 pm. Jaipur vendors only, three weeks to build. </hard_constraints> <quotas> Twenty themes. Eight safe enough for the first family meeting. Eight unusual but genuinely buildable here in February. Four I should only propose if the family turns out braver than they seem. </quotas> <task> An Artifact table: number, theme, one line of décor, one activity that involves the guests, Safe or Odd, and a column naming the constraint this theme is most likely to break - the budget, the cold, the five-year-olds or the eighty-five-year-olds. Fill that last column honestly for every row, including the safe ones. A theme that fails quietly at 9 pm in February is worse for me than a theme that fails obviously now. Do not rank them and do not pick three for me. Then, in your reply: which two of the twenty do you think a railwayman of that generation would actually find touching rather than clever? Say why in one line each. That is the one judgement I do want from you. </task>
Then loop it
- Themes 3 and 15 survive. For each, ten small touches that would surprise a man who spent his life on the railways - details his own daughter would not have thought of.
- Your four risky themes are not risky. Replace them with four I would be nervous to show the family, and tell me what each one risks.
- Mix theme 3 with the photo booth from 15. Put a rough décor list in the Artifact that fits ₹2.5 lakh with the lawn, the lighting and the cold accounted for.
Why it worksTwenty themes, safe and unusual mixed together, give Farhan raw ore to sift and combine instead of one tired answer.
Instead of
How can I increase sales in my bakery?
Claude version
<my_shop> A small bakery in Shillong. Busy mornings, busy weekends, and nothing at all between 2 and 5 pm on weekdays. Two colleges, a hospital and several offices within walking distance. Walk-in trade only - I have no delivery rider and no second oven. </my_shop> <quotas> Twenty-five ideas. Across four groups, each properly populated: new products, offers and combos, events inside the shop, tie-ups with the places nearby. At least five that I would not have arrived at myself - and a discount with a new name does not count as one of them. </quotas> <feasibility> If an idea needs something I have said I do not have - a rider, a licence, a bigger kitchen, staff after 6 pm - name that in the row. I would rather read it here than hear it from a supplier. </feasibility> <task> An Artifact table: number, idea in one line, group, cost to try (nothing, small, real money), who it is aimed at, and what it needs that I may not have. No ranking, no shortlist, no best three. Then, in your reply, two separate lists: the five ideas you think are obvious, and the five you nearly left out for being too odd. I want to see both ends of what you were weighing. </task>
Then loop it
- Ideas 7, 14 and 21 fit. For each, the first three steps I can take this week - who I speak to, what I buy, what goes on the board outside.
- Ten more in the tie-ups group alone, and include the odd ones this time. The colleges and the hospital are obvious; who else is on that road at three in the afternoon?
- Now a friendly WhatsApp message to my regulars announcing idea 14 - under 50 words, warm, and no discount language anywhere in it.
Why it worksAsking for 25 grouped ideas, with a quota of unconventional ones, turns one stubborn problem into a field of options.
Questions
What is The Seed of a New Idea in prompt engineering?
Do not ask for the perfect name. Ask for thirty rough ones. Volume and range are what the machine is genuinely better at than you - it will lay out more options in a minute than you would reach in a week - while the judgement stays where it belongs, with the person who knows the shop, the buyers and the budget. State a number, demand variety in named directions, forbid it from choosing a winner, and then do the choosing yourself.
When should I use this technique?
When you need a name, a theme, a headline or a way out of a stuck problem, and you can feel that your own first idea is also the most obvious one.
How do I apply it in Claude?
Put the quotas in a tag and they hold: so many in this direction, so many in that, so many that are deliberately risky. Ask for the output as a table in an Artifact and you get something you can actually work in - delete rows, keep rows, download the shortlist - instead of a list buried four messages back. One extra column does more than any other instruction: have it mark each idea obvious, decent or strange, by its own judgement. You will disagree with some of those marks, and the disagreement tells you where your own taste sits. When a whole direction is wrong, edit your first message rather than arguing downward from a bad list.
What are the limits of this technique in Claude?
Its instinct is taste, which in brainstorming is a handicap. Left alone it produces twenty things that are all mildly elegant and none of them odd, and when you show it a shortlist it will admire your choice and then generate more of the same, narrowing you exactly when you wanted widening. So make strangeness a quota rather than a request - five ideas you expect me to reject - and ask what it ruled out as too silly to include, which is often where the useful one was hiding. Long single lists also repeat themselves; ten at a time, with a named direction, stays sharp.
How do I improve the answer after the first try?
When the first list arrives, do not hunt for a single winner yet. Mark the three or four ideas that spark something, even if they are rough. Then loop: "Give me ten more like numbers 4 and 9," or "Combine 2 and 17." If the list feels too safe, say "Make the next ten bolder." If it drifts away from your situation, repeat one key fact, such as your budget or your customers. Two or three rounds usually turn ore into gold. Save the prompt that worked so you can reuse it next season.
Loop it
When the first list arrives, do not hunt for a single winner yet. Mark the three or four ideas that spark something, even if they are rough. Then loop: "Give me ten more like numbers 4 and 9," or "Combine 2 and 17." If the list feels too safe, say "Make the next ten bolder." If it drifts away from your situation, repeat one key fact, such as your budget or your customers. Two or three rounds usually turn ore into gold. Save the prompt that worked so you can reuse it next season.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.