Lesson 28 of 32

The Mirror of Self-Critique

Reflection Loops

Build the checking into the prompt itself. Give a short checklist of what the finished work must achieve - four to six things a careful reader could answer yes or no - then tell the model to draft, score itself against each point while quoting the line that falls short, and rewrite. Two rounds, not ten. It works because "make it better" has nothing to measure against, while five checkable criteria do.

Reach for it whenWhen you can say precisely what good looks like - a cover letter, a plan, a story for a particular class - and you would rather receive a draft that has already been through an editor than one you must edit yourself.

Write the rubric as a numbered list and name the number of rounds, because left to itself it will either skip the scoring or run forever. The real gain here is reuse: a rubric-driven reviewer makes an excellent Custom GPT - a strict cover-letter critic, a Class 3 reading checker - so that afterwards you paste only the draft and the standard comes with the tool. Run the rewrite in Canvas, where the scored version and your own edits sit in one document rather than scattered across six replies. With the reasoning models, do not bother telling it to think step by step; put your effort into the criteria instead, since those are the part it cannot infer.

From the book Save a rubric-driven reviewer as a GPT - for example, a strict cover-letter critic - so you only paste the draft each time. Ask it to open the final text in Canvas, where you can see the rewrite side by side and edit it yourself.
Worth knowingThe scoring is generous, and by round two almost everything comes back a five. Demand a named weakness in every round, and treat a perfect score as evidence that the mirror is flattering you rather than that the work is done. The deeper limit is this: a model marking its own paper is checking the draft against your criteria, not against the world. It cannot reliably notice that a figure, a date or a scheme name it produced is wrong, and it will happily score such a line as accurate - self-critique catches weak writing, not false facts. Anything checkable still needs web search or your own eyes afterwards.
  • Number the rubric and name the number of rounds, or it drifts.
  • Demand one real weakness per round; all fives means it flattered you.
  • Build a recurring critic as a Custom GPT; paste only the draft.
  • Use Canvas so the rewrite and your edits share one document.
  • Self-critique cannot catch a wrong fact. Verify those separately.
The Vague Polish vs. The Scored Rubric

Instead of

Improve my cover letter for this job.

ChatGPT version

Junior accounts assistant, logistics firm. B.Com 2025, Tally and Excel, six months interning with a CA firm in Bhubaneswar.

[Paste the job advertisement here.]

Write the letter. Then mark your own work in a table, one row per standard:

| # | Standard | Score /5 | The line that cost the mark |

The standards: word count below 200; the post named in the opening sentence; three or more phrases borrowed from the advertisement's own vocabulary; one specific thing I did at the CA firm rather than a description of where I sat; nothing a reader has met a thousand times before.

Rewrite after the first table, mark again, then stop. Two rounds.

Two rules for the marking. Every table must carry at least one score below 5, the last one included - straight fives tell me the marking has stopped working rather than that the letter is finished. And on the row about my internship, remember you have no way of knowing whether what you wrote there is true: if you supplied a detail I never gave you, write INVENTED in that row and ask me, rather than scoring it satisfied.

Open the final letter in Canvas so I can edit it myself.
Open ChatGPT 1,128 characters

Then loop it

  1. Round two came back all fives, which I do not believe. Be stricter: name one real weakness still in the final letter, quote it, and fix only that.
  2. You wrote that I "reconciled vendor ledgers monthly" - I did that twice, not monthly. That is the kind of error the rubric cannot catch. Correct it, then tell me which other sentences describe my experience in ways you could not have known.
  3. Good. Now give me the rubric by itself as a numbered list, written so I can save it as the instructions of a Custom GPT called "Strict letter critic" - I will paste a new advertisement each time I apply.

Why it worksThe rubric turns "make it better" into five checkable standards, and the scoring loop forces the AI to mend its own shortfalls before she sees the letter.

The Cheerful Plan vs. The Plan That Argued With Itself

Instead of

Should I grow vegetables instead of wheat?

ChatGPT version

I farm 10 acres of wheat and paddy near Moga, Punjab. I am thinking of moving 2 acres to tomato and capsicum next season to earn more.

Do this in three parts, clearly separated.

**Part 1 - the plan.** Steps through the season, where I would sell, and what I would need to learn. Keep it short.

**Part 2 - the attack.** Now argue against your own plan. The five strongest reasons it could fail for a farmer like me: price crashes, labour at harvest, water, storage, transport, anything else you see. Strongest first. No softening, no "however, with careful planning".

**Part 3 - the revision.** Rewrite the plan so that it answers each of the five objections, and say plainly where an objection cannot be answered and only accepted as a risk.

Then a short list of questions to put to the Krishi Vigyan Kendra before I decide anything.

One thing I want you to be honest about. Your objections will be sound in kind but you cannot know this season's prices, this year's input costs or what my local mandi is doing. Mark every number in Parts 2 and 3 that you are not in a position to know, and put those at the top of my KVK question list rather than inside the plan as if they were facts.
Open ChatGPT 1,192 characters

Then loop it

  1. Objection 2, the price crash, is the one that worries me. Play a hard-headed mandi trader in Moga and attack the revised plan again - on selling alone, not on cultivation. Do not be polite about it.
  2. Now something the critique cannot do: list every figure in this conversation that you produced from general knowledge rather than from my situation - yields, rates, costs - so I take all of them to the KVK instead of trusting any.
  3. Now give me the final plan as one page in simple Punjabi that I can put in front of my family, with the accepted risks kept in, not quietly dropped.

Why it worksMaking the AI argue against itself exposes risks a cheerful first answer hides, and the revised plan must survive those objections before he trusts it.

The Lone Writer vs. The Critic and the Creator

Instead of

Write a story for kids about saving water.

ChatGPT version

Two roles in one conversation, and I want to see both at work.

**Creator** - write a 250-word story for Class 3 children in Kerala about a girl who saves water at home. Simple English, a little mischief, a clean ending.

**Critic** - a primary-reading specialist, twenty years in front of Class 3, no interest in sparing the writer.

The Critic reports in a table, nothing else:

| Test | Pass or fail | The words at fault |

The tests: every word inside an 8-year-old reader's range; no sentence running past 12 words; the lesson made once rather than three times; the child arriving at the point instead of the story announcing it; one funny moment the class will repeat to each other afterwards.

Then the Creator rewrites in answer to the table. Then the Critic fills the table again. Two rounds and no more. Show me the final story and the last table, and leave the earlier drafts out of it.

A standing rule for the Critic: something fails in every table, the last one included. A table of passes means the Critic has stopped reading, and I would rather be told that.
Open ChatGPT 1,074 characters

Then loop it

  1. The Critic passed the preaching test it should have failed - the closing paragraph is a lecture. Critic: fill the table for the ending alone and quote the sentences. Creator: rewrite the ending and leave everything before it alone.
  2. Better. Now a check neither role can carry out: list the words in the final story an 8-year-old in Kerala might not have met in English, and mark where you are guessing rather than knowing. I read to this class every day and I am the one who can tell.
  3. Lovely. Add 5 simple comprehension questions for the class, and a Malayalam version of the story.

Why it worksSplitting creator and critic gives the self-check a strict, named judge and a rubric, so each rewrite answers real problems rather than vague praise.

Loop it

Read the scores before the story. If every criterion scores a perfect 5, the loop is flattering itself - reply, "Be stricter: quote the weakest line for each point and rewrite." If the final version improved one point but broke another, add that point to the rubric and run one more round. Stop after two or three rounds; more rarely helps. Then do your own human read, checking facts the rubric cannot catch. When a rubric works, save it and paste it into future prompts as a standing measure of quality.