The Mirror of Self-Critique

Tag the parts and the loop runs cleanly: <rubric> for the standard, <creator> and <critic> for the two characters, and the instruction to score, quote and rewrite placed last. Claude takes to this chapter more readily than the others because it will criticise its own text without being coaxed - ask what is weakest in a draft and you generally get a real answer rather than a compliment. Put the final version in an Artifact while the scores and the critic's notes stay in the chat, so the clean text is separate from the argument about it. A rubric you use often belongs in a Project's instructions, where every draft in that project is measured the same way.

From the book Put your favourite rubrics in a Project's instructions so every draft in that project is scored the same way. Ask Claude to show the final version in an Artifact while the scores and critique stay in the chat, so the clean text is easy to copy or download.

• <rubric>, <creator>, <critic> as tags; scoring instruction last.
• Ask what is weakest - Claude answers that honestly.
• Hold back your praise: it softens the critic for the whole chat.
• Final version in an Artifact, scores and notes in the chat.
• A rubric judges wording, not truth. Check facts yourself.

<job_advertisement>
[Paste the advertisement here, in full. Long material first, my instruction at the end.]
</job_advertisement>

<me>
B.Com, 2025. Tally and Excel. Six months interning at a CA firm in Bhubaneswar, where my actual work was [write the two or three things you really did]. Applying for junior accounts assistant at a logistics firm.
</me>

<rubric>
Answer each of these with yes or no before you put a mark out of 5 against it:
- Is the letter under 200 words?
- Does its first sentence say which post I am applying for?
- Does it pick up at least three phrases from the advertisement above?
- Does it contain one thing I actually did, as against where I happened to be?
- Is it free of the sentences every fresher writes?
</rubric>

<scoring_discipline>
Score each criterion out of 5. Quote the line responsible for any score below 5. Every round must name at least one genuine weakness, the final round included - a clean sheet of fives tells me the critic has gone to sleep, not that the letter is finished.

And be exact about the boundary of what you are checking: the rubric tests whether the letter is well made. It cannot test whether the letter is true. If you have written anything about my internship that I did not give you above, flag it under a heading <invented> and ask me, rather than scoring it as satisfied.
</scoring_discipline>

<task>
Write the letter, score it, rewrite it, score it again. Two rounds. Then put the final letter in an Artifact and leave the scores and your notes here in the chat.
</task>

<my_situation>
Ten acres of wheat and paddy near Moga. I am considering putting 2 acres into tomato and capsicum next season to raise my income. I have no experience of vegetables, and no cold storage of my own.
</my_situation>

<sequence>
First a plan: the season's steps, where I would sell, what I would have to learn. Brief.

Then turn on it. Give me the five strongest reasons this fails for a farmer in my position - price collapse, harvest labour, water, storage, transport, whatever you judge strongest. Order them by how likely they are to actually sink me. No "however, with good planning" anywhere in this section; the whole value of it is that it does not reassure me.

Then revise the plan so it answers each objection, and state openly which objections cannot be answered and must simply be carried as risk.

Then questions for the Krishi Vigyan Kendra.
</sequence>

<what_you_cannot_know>
Your objections will be right in kind. Your numbers will not be current - this season's rates, my mandi's behaviour, this year's input costs are all outside what you have. So mark every figure you produce with (unverified) and gather them at the end under <take_to_kvk>. Do not let an unverified number sit inside the revised plan as though it were established, because that is precisely the error self-critique is blind to: you will score the plan sound and the arithmetic underneath it may be two years stale.
</what_you_cannot_know>

<task>
Work through the sequence above.
</task>

<roles>
<creator>A writer of stories for small children, with energy and some mischief.</creator>
<critic>A primary-reading specialist who has taught Class 3 for twenty years and has no interest in sparing the writer's feelings.</critic>
</roles>

<brief>
A 250-word story for Class 3 children in Kerala, about a girl who saves water at home. Simple English, a little humour, a clear ending. I teach this class and will read it aloud.
</brief>

<rubric>
The Critic answers each of these before scoring anything:
- Which words here would stop an 8-year-old reader?
- Which sentences run past twelve words?
- Is the lesson made once, or does it keep returning?
- Does the child in the story work the point out, or is the point announced to the reader?
- Where is the moment the class will repeat to each other at lunch? If there is none, say so.
</rubric>

<how_the_two_work>
Each fault must come with the exact words attached. A verdict with nothing quoted beneath it is no use to someone who has to read this out tomorrow.

If you reach a pass where you genuinely cannot fault anything, do not manufacture a fault to satisfy me. Say instead which of the five questions you are least confident you applied properly, and why - that is the more useful admission, and it is the one a self-marking loop almost never makes.

I am withholding my own praise until the whole thing is finished. Approval softens a critic, and I would rather keep the critic.
</how_the_two_work>

<task>
Two passes. Final story in an Artifact; the notes stay down here; the discarded drafts go nowhere.
</task>