Lesson 7 of 32 · Part One, the principles
The Virtue of Persistent Effort
Iteration
Your first prompt is a hypothesis, not a finished article. Iteration means reading the answer like a detective, naming one thing that is wrong - the prices, the tone, the length - and fixing only that in your next message, so each round keeps what already worked and the good parts survive.
How to use The Virtue of Persistent Effort in Claude
Two mechanics make the loop cleaner here. An Artifact gives the work one body: ask for the menu or the table as an Artifact and every round edits that same document in place, so you are never comparing draft two with draft five by scrolling. And editing an earlier message makes Claude answer that version afresh, which is how you abandon a wrong turn instead of arguing your way out of it - far cleaner than piling another correction on top. Structure the first prompt with tags so later rounds have something precise to point at: <current_prices> and <must_include> are addressable, a paragraph is not. Then add one unusual instruction that pays for itself - ask it to tell you if a change you requested has made the thing worse.
- Keep the work in one Artifact and refine it in place.
- Edit an earlier message to abandon a wrong turn cleanly.
- Ask each round: "has this change made it worse?"
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Two mechanics make the loop cleaner here. An Artifact gives the work one body: ask for the menu or the table as an Artifact and every round edits that same document in place, so you are never comparing draft two with draft five by scrolling. And editing an earlier message makes Claude answer that version afresh, which is how you abandon a wrong turn instead of arguing your way out of it - far cleaner than piling another correction on top. Structure the first prompt with tags so later rounds have something precise to point at: <current_prices> and <must_include> are addressable, a paragraph is not. Then add one unusual instruction that pays for itself - ask it to tell you if a change you requested has made the thing worse.
From the book If a whole chat went wrong, edit your earlier message and try a different version instead of starting over. Keep the final document as an Artifact so each refinement updates the same panel.
- Keep the work in one Artifact and refine it in place.
- Edit an earlier message to abandon a wrong turn cleanly.
- Ask each round: "has this change made it worse?"
- Tag the parts you will point at later: <prices>, <must_include>.
- When rounds circle, ask for one improved prompt and restart.
Instead of
Make a Diwali offer list for my parlour.
Claude version
<context> I run a small beauty parlour in a residential area of Indore. My clients are mostly working women and college girls. </context> <current_prices> Facial ₹800 Threading ₹60 Manicure ₹400 Hair spa ₹900 </current_prices> <what_i_want> A Diwali offer menu for WhatsApp. Five packages, built by combining the services above - threading plus facial, a bridal trial, and so on. Each package: name, what is included, normal total, offer price in ₹. Every discount between 10 and 15 percent, with the arithmetic shown beside it so I can check your sums. Friendly and festive, the way I would talk to a regular. </what_i_want> <how_we_will_work> This is round one of several. Build it as an Artifact so each correction edits the same menu rather than producing a new one. If a change I ask for later would damage the menu, say so before you make it. </how_we_will_work> <task> First, three or four lines of your own thinking on which combinations will actually sell to college girls and which to working women - I want to argue with your reasoning before I accept the prices. Then the menu. </task>
Then loop it
- Your reasoning on the bridal trial is wrong for my area - very few bridal clients here. Edit the Artifact: swap that package for something aimed at college girls, and recalculate every discount so each sits between 10 and 15 percent.
- Now rename all five packages with short, catchy Hindi names and add a closing line about booking on WhatsApp. If a Hindi name makes any package unclear, tell me rather than keeping it.
- Shorten the whole Artifact so it fits one phone screen without scrolling. Then rewrite my original prompt, tags and all, with blanks for [festival] and [current prices], so I can save it in a Project called "Parlour offers" and run it for New Year.
Why it worksEach follow-up fixes one flaw the first answer revealed - price, name, length - until the menu is ready to send, and the last step turns it into a reusable format.
Instead of
Write about my Pochampally saree.
Claude version
<about_me> I am a third-generation ikat weaver in Pochampally, Telangana. I sell cotton and silk sarees online. </about_me> <must_include> The saree is handwoven. The tie-and-dye pattern is planned thread by thread before the weaving starts. One saree takes about [number] days. </must_include> <voice> Warm, as though the weaver is speaking to the buyer. Not a brand. Not an advertisement. </voice> <task> Write a story card of about 60 words to tuck inside each saree package. Give me three attempts that differ in approach rather than in wording, labelled A, B and C, and one line under each saying what it is trying to do. I will choose one, and we will refine only that one, as an Artifact. </task>
Then loop it
- A is the one, but it still sounds like an advertisement. Put A into an Artifact and make it humbler - a thank-you note from my family rather than a claim about quality. Tell me if losing the claims makes the card feel thin.
- Better. In the same Artifact, give me three versions holding that voice: silk, cotton, and one for someone buying it as a gift.
- Now add a Telugu translation of the cotton version to the Artifact, and list the lines that may sound unnatural in Telugu so I can check them with my daughter.
Why it worksThe weaver treats the first draft as feedback, steering tone, then variety, then language, instead of giving up when the first card sounds like an advertisement.
Instead of
How to save money as a cab driver?
Claude version
<context> I drive a cab full-time in Bengaluru through ride apps, six days a week. </context> <what_i_need> A weekly table I fill in each night. Columns: date, hours driven, trips, total earnings, fuel or CNG cost, app commission, money taken home. Under the table, five short points on how to read it at the end of the week to find my best and worst days. Simple English throughout. </what_i_need> <how_we_will_work> Make it an Artifact. I will add columns, then paste a real week's figures into it, so I want one sheet that keeps improving rather than a fresh one each round. </how_we_will_work> <task> Build the table and the five points. Before them, say in two lines which single column will tell me the most about my week and why. </task>
Then loop it
- Add two columns to the Artifact: toll and parking, and small expenses like tea and food. Keep the existing columns and the five points untouched. Say if seven columns is now too many to fill in at night.
- Here is last week's data: [paste numbers]. Fill the Artifact table and tell me which day, and which hours of the day, gave the best take-home per hour. Show how you worked that out.
- Good. Turn the five points into a Sunday checklist inside the same Artifact, with a simple Kannada version below it.
Why it worksThe loop moves from an empty template to his real numbers to a weekly habit, each round built on what the previous answer got right or missed.
Questions
What is The Virtue of Persistent Effort in prompt engineering?
Your first prompt is a hypothesis, not a finished article. Iteration means reading the answer like a detective, naming one thing that is wrong - the prices, the tone, the length - and fixing only that in your next message, so each round keeps what already worked and the good parts survive.
When should I use this technique?
When the first answer is nearly right but wrong in one way you can name, and when the thing you are making is something you will send, print or reuse rather than glance at.
How do I apply it in Claude?
Two mechanics make the loop cleaner here. An Artifact gives the work one body: ask for the menu or the table as an Artifact and every round edits that same document in place, so you are never comparing draft two with draft five by scrolling. And editing an earlier message makes Claude answer that version afresh, which is how you abandon a wrong turn instead of arguing your way out of it - far cleaner than piling another correction on top. Structure the first prompt with tags so later rounds have something precise to point at: <current_prices> and <must_include> are addressable, a paragraph is not. Then add one unusual instruction that pays for itself - ask it to tell you if a change you requested has made the thing worse.
What are the limits of this technique in Claude?
Claude is obliging about refinements, which sounds like a virtue and is not. Ask for shorter and you get shorter, even when shorter has gutted the thing; you will not be warned, and the answer stays confident. That is why "tell me if this change has damaged it" belongs in every round. The second trap is the long chat: Claude holds a great deal of context, so an abandoned draft from twenty messages back is still pulling. When the rounds start circling, ask it to rewrite your requirements as one improved prompt and begin again in a clean chat with the Artifact's final text pasted in.
How do I improve the answer after the first try?
Iteration follows this book's five-step loop: Aim, Prompt, Read, Refine, Reuse. After the first answer, read it like a detective and name one specific flaw - too long, wrong tone, a price that is off. Fix only that in your next message: "Keep everything but make the discounts 10 to 15 percent." Fixing one thing at a time keeps the good parts. When the answer is finally right, ask the AI to rewrite your original prompt with all the fixes included, and save that improved prompt for next time.
Loop it
Iteration follows this book's five-step loop: Aim, Prompt, Read, Refine, Reuse. After the first answer, read it like a detective and name one specific flaw - too long, wrong tone, a price that is off. Fix only that in your next message: "Keep everything but make the discounts 10 to 15 percent." Fixing one thing at a time keeps the good parts. When the answer is finally right, ask the AI to rewrite your original prompt with all the fixes included, and save that improved prompt for next time.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.