Lesson 23 of 32 · Part One, the principles
The Reversal of Perspective
When an answer comes back strangely off-target, the usual reaction is to type the same request again, louder. The better move is to turn the question around: ask the machine to tell you, in its own words, what it thinks you asked for. In the gap between its account and your intention you will find the exact word that went wrong - "clean up", "young adults", "about water" - and one word fixed is worth ten more attempts.
How to use The Reversal of Perspective in Claude
Put the diagnosis in tags and it stops sliding back into the task: <my_earlier_request> holds what you actually sent, <what_i_got> says what came back, and <task> asks for the mirror - your reading of my goal, the assumptions you made, and the words that caused them. Add a <do_not> block forbidding the deliverable, because the instinct to be helpful is strong and a half-paraphrase followed by another wrong attempt teaches you nothing. When the gap is clear, edit your original message with the fix and re-run it, so the corrected request stands in place of the confused one. If the same misunderstanding keeps returning across chats, the fix belongs in the Project's standing instructions, not in your memory.
- Separate tags for what you sent, what came back, and the diagnosis.
- A <do_not> block: no apology, no revised version, no code.
- Insist on quoted words, not "I misunderstood your intent".
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Put the diagnosis in tags and it stops sliding back into the task: <my_earlier_request> holds what you actually sent, <what_i_got> says what came back, and <task> asks for the mirror - your reading of my goal, the assumptions you made, and the words that caused them. Add a <do_not> block forbidding the deliverable, because the instinct to be helpful is strong and a half-paraphrase followed by another wrong attempt teaches you nothing. When the gap is clear, edit your original message with the fix and re-run it, so the corrected request stands in place of the confused one. If the same misunderstanding keeps returning across chats, the fix belongs in the Project's standing instructions, not in your memory.
From the book Once Claude's restatement shows the gap, edit your original message with the fix and re-run it, so the corrected version replaces the confused one. If the same misunderstanding keeps coming back, put the clarification in your Project's standing instructions.
- Separate tags for what you sent, what came back, and the diagnosis.
- A <do_not> block: no apology, no revised version, no code.
- Insist on quoted words, not "I misunderstood your intent".
- Re-run the corrected original rather than stacking fixes.
- A recurring misreading belongs in the Project's instructions.
Instead of
Clean up this customer table for me.
Claude version
<my_earlier_request> Clean up this customer table for me. </my_earlier_request> <what_you_produced> Code that deleted rows from the table. </what_you_produced> <task> Diagnose the gap between those two. Answer in four numbered parts: 1. The goal you took me to have. 2. The meaning you gave "clean up" - and the exact words of mine that led you to it, quoted. 3. Every assumption you made about the schema, the columns, the duplicates, and what could safely be removed. 4. What you need from me before any statement is written. </task> <do_not> Write code. Apologise. Offer a corrected version. Explain what you will do better next time. </do_not>
Then loop it
- I see the gap - I meant fix formats for phone numbers and dates, not delete duplicates. Restate the task again with this correction before writing any code.
- Correct. Now write the SQL for PostgreSQL, with a comment above each step.
Why it worksAsking the AI to explain its reading of "clean up" exposes the exact word that caused the wrong result.
Instead of
Plan a reading programme for young adults.
Claude version
<situation> I am a librarian at a public library in Kolkata. I asked you to plan a reading programme for young adults and you returned a list of heavy college texts. Something in my wording sent you there. </situation> <task> Paraphrase my earlier request in two or three sentences. Then set out, as a short list: - who you pictured as the readers; - the age range you assigned to "young adults"; - the language you assumed the books would be in; - the kind of events you thought I wanted. Against each one, quote the phrase of mine that produced it. </task> <do_not> Suggest a single new book, event or plan in this reply. </do_not>
Then loop it
- Now I understand - you took young adults to mean 18 to 25. I meant teenagers aged 13 to 17, reading in Bengali and English. Restate my request once more, then give the plan.
- Good. Now write short poster text for the notice board, in Bengali.
Why it worksThe paraphrase reveals that a common phrase meant something different to the AI than it did to her.
Instead of
Write a notice about water for the village.
Claude version
<my_earlier_request> Write a notice about water for the village. </my_earlier_request> <what_you_produced> An essay on saving water. </what_you_produced> <task> Tell me three things, and nothing else: 1. What work you think that text was meant to do. 2. Who you imagined reading it, and where you imagined it being displayed. 3. Which words of mine pushed you towards an essay. Quote them. Then write out the request I should have sent, as a prompt I can copy - not as advice about being more specific. </task> <context>I am the sarpanch of a village near Nashik. Whatever I put up has to be read standing in front of a board, by people on their way to work.</context>
Then loop it
- I see - "about water" was too wide. It is a notice for the gram panchayat board about new tanker timings: Monday, Wednesday and Friday, 7 to 9 am, near the school. Write it in Marathi, under 80 words.
- Add one line asking families to keep their containers ready before the tanker arrives.
Why it worksBy asking the AI to point to the words that misled it, Sangita learns how to write a clearer request next time too.
Questions
What is The Reversal of Perspective in prompt engineering?
When an answer comes back strangely off-target, the usual reaction is to type the same request again, louder. The better move is to turn the question around: ask the machine to tell you, in its own words, what it thinks you asked for. In the gap between its account and your intention you will find the exact word that went wrong - "clean up", "young adults", "about water" - and one word fixed is worth ten more attempts.
When should I use this technique?
When the output is not merely imperfect but aimed somewhere else entirely, and you cannot see why - before you rewrite the prompt, make it show you what it heard.
How do I apply it in Claude?
Put the diagnosis in tags and it stops sliding back into the task: <my_earlier_request> holds what you actually sent, <what_i_got> says what came back, and <task> asks for the mirror - your reading of my goal, the assumptions you made, and the words that caused them. Add a <do_not> block forbidding the deliverable, because the instinct to be helpful is strong and a half-paraphrase followed by another wrong attempt teaches you nothing. When the gap is clear, edit your original message with the fix and re-run it, so the corrected request stands in place of the confused one. If the same misunderstanding keeps returning across chats, the fix belongs in the Project's standing instructions, not in your memory.
What are the limits of this technique in Claude?
Claude's instinct when told it got something wrong is to apologise and correct course immediately, which is precisely what you do not want here - an apology plus a better attempt hides the diagnosis you asked for. Say in as many words: no apology, no revised version, diagnosis only. The second habit to guard against is self-blame that explains nothing. "I misunderstood your intent" is not information. Insist that it quote the exact words from your prompt that produced each assumption; quoted words can be fixed, while a confession cannot.
How do I improve the answer after the first try?
Use this principle when an answer is strangely off-target. Do not retype the same request louder. First ask, "What do you think I asked for?" Read the reply and look for the gap: a word read differently, a wrong assumption about who the reader is, or a main goal buried under details. Fix that one thing in your next message, ask the AI to restate the task once more, and only then let it produce the answer. Keep the corrected prompt for next time.
Loop it
Use this principle when an answer is strangely off-target. Do not retype the same request louder. First ask, "What do you think I asked for?" Read the reply and look for the gap: a word read differently, a wrong assumption about who the reader is, or a main goal buried under details. Fix that one thing in your next message, ask the AI to restate the task once more, and only then let it produce the answer. Keep the corrected 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.