The Virtue of Persistent Effort (Iteration)

Set the iteration method out in a <working_method> block before the task, so the rounds are agreed in advance rather than negotiated turn by turn. Artifacts are where the evolving draft lives - the version updates in place and the chat stays for discussion of it. Claude also responds well to being asked what it thinks is weak in its own output, which turns a round of your guessing into a round of its diagnosis.

• Declare the rounds up front in a <working_method> block.
• Keep the evolving draft in an Artifact, discussion in chat.
• Ask it to name its own weakest sentence each round.
• Say what to do when your feedback is ambiguous: ask, don't guess.
• Put recurring background in a Project, not every prompt.

<context>
I am writing noir fiction. Setting: a rain-soaked, neon-lit Neo-Hyderabad in 2077. Protagonist: Kaito, a cynical, world-weary private detective. The opening scene: Kaito in his small office overlooking a crowded market when a mysterious client walks in holding a data chip containing an impossible photograph.
</context>

<working_method>
We are iterating. I expect several rounds, and I want each round to be narrow.

Round 1: options only, no prose.
Round 2: draft the chosen option into an Artifact.
Round 3 onwards: targeted revision of named passages inside that same Artifact.

When I say a passage is not working, do not immediately rewrite it. First tell me what you think is wrong with it in one line, then offer two alternatives. If you genuinely cannot tell what I am objecting to, ask - a wrong guess costs us a whole round.
</working_method>

<task>
Start with Round 1 only.

Give me three possible openings. For each, give exactly: the first line, the mood in three words, and the mystery the impossible photograph would turn out to conceal. Under 40 words each.

Then, separately, one line on which of the three you think is strongest as noir and why. I want your judgement, not neutrality - I will disagree if I disagree.

Do not write the scene yet.
</task>

<context>
I run 'Deccan Brews', an independent coffee shop in Kondapur, Hyderabad. My customers are mostly IT professionals from nearby offices. My marketing budget is ₹20,000 per month.
</context>

<working_method>
Three rounds. Do not skip ahead.

Round 1 - interrogate the brief. What do you need to know before a 4-week plan would be worth anything? Up to eight questions, then stop.
Round 2 - build the plan as an Artifact so we can edit it across turns without you reprinting it each time.
Round 3 - attack your own plan.

Put the Deccan Brews background and this working method into a Project so I do not restate it in every chat about the shop.
</working_method>

<task>
Round 1 only.

Ask me your questions. Cover at minimum: footfall and peak hours, current average bill, repeat-customer share, seating and dwell time, delivery-app presence, and the nearest competing cafes.

Before the questions, give me one short paragraph headed "What I am assuming" - the assumptions you are currently making about a Kondapur IT-corridor cafe. I would rather correct a stated assumption than discover a hidden one in Round 2.

If any question is one you expect a small independent owner not to have the data for, mark it and say what cheap proxy would do instead.
</task>

<context>
I need to explain the attention mechanism in a transformer model to business managers with no technical background. I will be presenting it. The analogy I want to build on: a student deciding which sentences in a textbook matter for answering a specific question, and how the model's scores on words parallel the student's choices.
</context>

<constraints>
- Under 120 words in the explanation itself.
- No jargon whatsoever: not 'vector', not 'weights', not 'token', not 'embedding'.
- The analogy must do the explaining. Do not append a technical definition to rescue it.
</constraints>

<working_method>
Three rounds, narrow each time. I will name the sentence that loses me; rewrite only that sentence, three ways, and let me choose. Do not rewrite the whole piece because one line was weak. Keep the live version in an Artifact.
</working_method>

<task>
Round 1.

First, write the explanation under <explanation>.

Then, under <where_this_lies>, tell me the two most important things the student-and-textbook analogy gets wrong about real attention, and for each one say whether a business manager actually needs to know it or whether it is a safe simplification. Be honest here even if it undermines the analogy - I would rather find out now than from someone in the audience.

Finally, under <riskiest_sentence>, name the one sentence you expect me to push back on.

If I later ask you to simplify something that is already as simple as it can be without becoming false, tell me so rather than complying.
</task>