The Virtue of Persistent Effort (Iteration)

Canvas is the iteration surface: it holds a document or code side by side with the chat, so you can highlight a passage and ask for that passage to be changed instead of asking for the whole thing again. Open it as soon as you know you will revise more than once. Structure the first prompt as a sequence of rounds - options, then draft, then targeted edits - so the model knows not to produce a finished artefact on turn one.

• Open a Canvas before the second revision, not after the fifth.
• Ask for options first; prose only once you have chosen.
• Say 'change only the highlighted text' on every edit.
• Put stable background into Custom Instructions, not each prompt.
• Ask for a diagnostic question before any rewrite.

Open a Canvas for this - we are going to work on it across several turns, and I want to edit the text directly rather than re-reading the whole story each time.

TURN 1 - I only want options, not prose.
Act as a gritty noir novelist. Before writing anything, give me three different openings for a story about Kaito, a cynical private detective in a rain-soaked, neon-lit Neo-Hyderabad in 2077. For each option give me only: the first line, the mood in three words, and what the central mystery would turn out to be. Keep each under 40 words. Do not write the scene yet.

I will pick one. Then:

TURN 2 - write the opening two paragraphs of the chosen option into the Canvas. Kaito in his small office above a crowded market; a client arrives with a data chip holding an impossible photograph.

TURN 3 onwards - I will highlight specific passages in the Canvas and ask for those to be changed. When I do, change only the highlighted text and leave the rest exactly as it stands. Do not silently improve other sentences.

One standing rule for all turns: when I say something is not working, ask me one diagnostic question before rewriting. Guessing at what I meant and producing a fresh draft wastes both our turns.

I run 'Deccan Brews', an independent coffee shop in Kondapur, Hyderabad. Main customers are IT professionals from the nearby offices. Marketing budget is ₹20,000 per month.

Do not give me the plan yet. We are going to build it in three passes.

PASS 1 - Ask me questions. What do you need to know about footfall, peak hours, current average bill, seating, delivery-app presence and the competition nearby before a 4-week plan would be worth anything? Ask up to eight questions and stop there.

PASS 2 - After I answer, open a Canvas and draft the plan as a table: Week, Action, Channel, Estimated Cost, Key Metric. One specific low-cost action per week, total staying inside ₹20,000.

PASS 3 - Then stress-test your own plan in the Canvas comments, not by rewriting it: which week is weakest, which cost estimate is least reliable, and what would have to be true for this plan to fail.

Across all passes: if an action needs something I have not told you I have - an Instagram account with reach, a POS that tracks repeat customers - say so explicitly rather than assuming it.

If my background and budget are going to come up in every marketing conversation, tell me what to put into Custom Instructions or saved memory so I stop retyping it.

I want to understand the attention mechanism in a transformer model. My audience is business managers with no technical background - I will be presenting this.

Build it with me in three rounds rather than handing me a finished explanation.

ROUND 1 - Explain it in under 120 words using one analogy: a student deciding which sentences in a textbook matter for answering a specific question. Show how the model's 'scores' map onto the student's choices. No jargon at all - not 'vector', not 'weights', not 'token'.

ROUND 2 - I will tell you which sentence lost me. Rewrite only that sentence, three different ways, and let me pick. Do not rewrite the whole explanation.

ROUND 3 - Once the wording holds, tell me where the analogy breaks down - the two things about real attention it gets wrong - and whether a business manager needs to know that or not.

Then put the final version into a Canvas so I can shape it into speaker notes.

Standing rule: if I ask you to simplify something that is already as simple as it can honestly get, tell me that instead of simplifying it into something false.