Lesson 21 of 32 · Part One, the principles
The Master's Review
Two reviews make the difference between a gamble and a plan. The first is yours: before you press send, read your own prompt as if you were the machine receiving it, and look for the word that could be read two ways or the detail you forgot. The second you ask for: have the model inspect your request for gaps before it answers, and inspect its own draft afterwards - through the eyes of the person who will read it.
How to use The Master's Review in Claude
Tags let you hand over the review criteria as a separate instrument from the task, which is the whole move in this chapter: <review_criteria> holds the three things you want looked at, and <task> says inspect my request first, then answer. Put a <before_you_answer> block at the top and the gap-finding happens before any words are wasted. Two app mechanics help. When its questions expose a hole in your prompt, edit that earlier message and re-run it, so the corrected request replaces the muddled one instead of trailing behind it. And for a long piece like a memoir chapter, ask for the revised text as an Artifact, which you can compare side by side with your own and download.
- Put the review criteria in their own tag, apart from the task.
- Open with <before_you_answer> so gaps surface before the draft.
- Edit the earlier message to re-run a corrected prompt.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Tags let you hand over the review criteria as a separate instrument from the task, which is the whole move in this chapter: <review_criteria> holds the three things you want looked at, and <task> says inspect my request first, then answer. Put a <before_you_answer> block at the top and the gap-finding happens before any words are wasted. Two app mechanics help. When its questions expose a hole in your prompt, edit that earlier message and re-run it, so the corrected request replaces the muddled one instead of trailing behind it. And for a long piece like a memoir chapter, ask for the revised text as an Artifact, which you can compare side by side with your own and download.
From the book Spotted a flaw in your prompt after sending it? Edit your earlier message and re-run it instead of piling on corrections. For long pieces like a memoir chapter, ask for the revised text as an Artifact so you can compare and download it.
- Put the review criteria in their own tag, apart from the task.
- Open with <before_you_answer> so gaps surface before the draft.
- Edit the earlier message to re-run a corrected prompt.
- Paste the provision; self-review checks prose, not statute.
- When you overrule a critique, ask whether your objection holds.
Instead of
Explain GST input tax credit for my exam.
Claude version
<who_i_am>CA Intermediate student, Surat. Exam in May.</who_i_am> <my_request> One page of revision notes on input tax credit under GST: the conditions for claiming it, blocked credits under Section 17(5), and the time limits. Short points rather than paragraphs, plus one worked example with round figures. </my_request> <before_you_answer> Read <my_request> back as though you were receiving it cold. Name anything ambiguous or missing, and ask me up to three questions. Then stop - do not begin the notes in the same message. </before_you_answer> <after_you_write> Close with a block headed "Not verified". In it, list every section number, rate, percentage and time limit you have produced from memory rather than from something I gave you. Include the ones you are confident about: confidence is not a source, and I would rather check twelve items than be caught by one. </after_you_write>
Then loop it
- Good questions. My exam is in May, I know the basics already, so give most space to blocked credits.
- Now quiz me with five tricky MCQs from these notes. Put the answers at the end.
Why it worksIt makes the AI inspect the order for gaps before acting, and review its own answer for weak points afterwards.
Instead of
Is my chapter good? Please check it.
Claude version
<chapter> [Paste the full 1,200 words here - the long material goes first, my instruction comes last.] </chapter> <about_it> The opening chapter of my memoir. I am a retired schoolteacher, and this is about my first posting, at a village school near Kottayam in 1978. </about_it> <role>A kind but honest memoir editor. Not a copy-editor, not a cheerleader.</role> <review_criteria> 1. Points where a reader loses hold of the chronology, or of which person is which. 2. Sentences that have outgrown themselves. 3. Moments where a feeling is announced rather than shown. </review_criteria> <do_not> Rewrite the chapter. Smooth my sentences. Change my plain style into a literary one. </do_not> <task> Give up to ten numbered suggestions against the criteria above, each quoting the line it refers to. Where you are unsure whether something is a flaw or simply my voice, say so and leave the decision to me. </task>
Then loop it
- Thank you. I disagree with suggestion 4 - that headmaster really was that strict. Skip it, and show me how you would fix suggestion 2 only.
- Read your own suggestions again. Which three matter most for a first-time reader?
Why it worksA named set of review criteria turns a vague "check it" into a focused critique she can act on line by line.
Instead of
Write a sorry message to an angry client.
Claude version
<situation> I photograph weddings in Jaipur. A couple's album is three weeks late; the printing lab's machine broke down. They are angry, and they are right to be. </situation> <must_contain> An apology that does not shelter behind the lab. The new delivery date: 20 November. Ten extra prints, free. </must_contain> <limits>WhatsApp message, under 100 words, in English.</limits> <task> Two blocks. First, the draft. Second, headed "As the bride would read it": review your own draft in her voice, on her phone, still angry. Quote any line that reads as defensive, cold or templated, and say what she thinks when she reaches it. Then one improved version - one, not a choice of three. Do not tell me the draft is strong. If it needs replacing, replace it. </task>
Then loop it
- Better. The second version still says "unfortunately" twice. Remove that and make it warmer.
- Now give me a Hindi version I can send to the bride's mother.
Why it worksAsking the AI to critique its draft through the reader's eyes catches tone problems before the message is sent.
Questions
What is The Master's Review in prompt engineering?
Two reviews make the difference between a gamble and a plan. The first is yours: before you press send, read your own prompt as if you were the machine receiving it, and look for the word that could be read two ways or the detail you forgot. The second you ask for: have the model inspect your request for gaps before it answers, and inspect its own draft afterwards - through the eyes of the person who will read it.
When should I use this technique?
Before anything that goes out of your hands - a message to an upset customer, notes you will revise from, a chapter you are about to send to a reader - and whenever your first instinct is to type "is this good?"
How do I apply it in Claude?
Tags let you hand over the review criteria as a separate instrument from the task, which is the whole move in this chapter: <review_criteria> holds the three things you want looked at, and <task> says inspect my request first, then answer. Put a <before_you_answer> block at the top and the gap-finding happens before any words are wasted. Two app mechanics help. When its questions expose a hole in your prompt, edit that earlier message and re-run it, so the corrected request replaces the muddled one instead of trailing behind it. And for a long piece like a memoir chapter, ask for the revised text as an Artifact, which you can compare side by side with your own and download.
What are the limits of this technique in Claude?
Claude's weakness here is agreeableness, and it runs in both directions. Tell it the headmaster really was that strict and it will drop the suggestion without argument, even when the suggestion was right - so when you overrule a critique, ask whether your objection actually holds. In the other direction, asked to review its own draft it will find real problems of tone and rhythm and almost never say "this fact may be wrong", because it cannot tell a remembered section number from a verified one. On anything legal or statutory, the honest arrangement is to paste the provision into the prompt and tell it to work only from that - otherwise its self-review is a review of its prose, not of its accuracy.
How do I improve the answer after the first try?
With this principle, the loop begins before you even press Send. Read your prompt once as if you were the AI: is anything ambiguous or missing? Then, when the first answer arrives, ask the AI to review it - "What is weak or uncertain in your answer?" or "Check this against my original request." Read its critique, accept what is useful, overrule what is not, and ask for a revised version. A single review round usually lifts the quality more than any other follow-up.
Loop it
With this principle, the loop begins before you even press Send. Read your prompt once as if you were the AI: is anything ambiguous or missing? Then, when the first answer arrives, ask the AI to review it - "What is weak or uncertain in your answer?" or "Check this against my original request." Read its critique, accept what is useful, overrule what is not, and ask for a revised version. A single review round usually lifts the quality more than any other follow-up.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.