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.
The chapter works here much as written, and the thing worth building is permanence: a Gem holding your rubric and the critic's character turns the standard into something standing, so you paste a draft and it is marked without being asked. What Gemini adds beyond the other two is a partial answer to the chapter's own warning. The critic cannot judge facts, but after the loop ends, "double-check response" will compare the statements in the final version against Search and show you which ones it can and cannot support - which is the nearest thing any of these tools offers to an external check on a self-marked draft.
From the book Build a Gem with your rubric and the critic's role written into its standing instructions. For factual claims the critic cannot judge, use "double-check response" to compare the answer against Google Search.
- Put the rubric and the critic's character into a Gem's instructions.
- Write "name one real weakness every round" into the Gem itself.
- Run double-check after the loop, on the final version only.
- A highlight is a flag, not proof - open the source.
- Claims about your own life cannot be double-checked. Read those yourself.
Instead of
Improve my cover letter for this job.
Gemini version
Junior accounts assistant, logistics firm, Bhubaneswar. I finished B.Com in 2025, I work in Tally and Excel, and I interned six months with a CA firm. [Attach the job advertisement - a screenshot or the PDF is fine - rather than retyping it.] Read the attached advertisement first and tell me in three lines what it is really asking for, as against what it lists. Then write the letter. Then mark your own letter on five counts, out of 5 each, quoting whatever lost the mark: its length against a 200-word ceiling, whether the opening sentence names the post, how closely it borrows the advertisement's vocabulary, whether there is one real piece of work in it rather than a job description, and whether any sentence in it could have been written by any other applicant. Rewrite, mark again, stop after the second round. Every round names one real weakness. If a round reports nothing wrong, say so plainly and I will assume the marking has gone slack. When the final letter is settled, do two things. Run double-check on it and tell me which statements Search can support - I expect it to confirm almost nothing, because this letter is about me, and I want to see that limit for myself. Then export the letter to Google Docs.
Then loop it
- Round two gave everything 5 out of 5. Name the weakest line that survives and quote it, then mend that alone.
- Double-check could not touch the internship sentences, as expected. So go through them with me: list each claim you made about my work there, and I will tell you which are true. One is not - "reconciled vendor ledgers monthly" should be twice in six months.
- Corrected. Now write the whole thing up as standing instructions for a Gem called "Accounts letter critic" - rubric, the one-weakness rule, and the instruction to read an attached advertisement first - so next time I attach the ad and nothing else.
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.
Instead of
Should I grow vegetables instead of wheat?
Gemini version
Ten acres of wheat and paddy near Moga. I want to shift 2 acres to tomato and capsicum next season for better income. No vegetable experience, no cold store. Lay out a short plan first - the season's steps, where I would sell, what I must learn. Then attack it. Five reasons this fails for a farmer like me, strongest first, and do not cushion any of them. Prices, labour at picking time, water, storage, the trip to the mandi - whatever you judge most dangerous. I am not looking for encouragement; I have enough of that from the man selling seed. Then revise the plan to meet each objection, saying where an objection can only be accepted and not solved. Now the part that needs Search rather than reasoning. Ground these specifically and give me the link and the date beside each: recent tomato and capsicum price behaviour in Punjab mandis, what Agmarknet or the state marketing board currently shows; whether a glut or crash has been reported in the last two seasons; what PAU or the Krishi Vigyan Kendra currently advises for vegetables in the Moga belt; and whether any scheme exists for small cold storage or crates. Where you cannot find a current figure, say so rather than offering one from memory. Finish with the questions I should put to the KVK, with the unsourced numbers at the top of that list.
Then loop it
- Prices are the real danger. Run double-check on your price section and show me which claims Search supports and which it cannot. I would rather see the gaps than a tidy paragraph.
- Now play a hard mandi trader from Moga and attack the revised plan on selling alone - grading, timing, who really sets the rate in a glut week. Do not search for this round; I want the argument, not the data.
- Write the final plan as one page in simple Punjabi for my family, keeping the accepted risks in plain sight, and export it to Google Docs so I can print it for our discussion.
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.
Instead of
Write a story for kids about saving water.
Gemini version
Build this as a Gem I will reuse all year, called "Class 3 story checker". Here is what belongs in its standing instructions, and I want you to run it once now on a water story. Two characters. The Creator writes stories for Class 3 children in Kerala - simple English, a little mischief, a clean ending, around 250 words. The Critic is a primary-reading specialist of twenty years' standing who marks without kindness. The Critic's standing checklist: vocabulary an 8-year-old reader can carry; no sentence over a dozen words; a single lesson rather than a repeated one; a point the child reaches instead of one delivered to the reader; and one moment of real fun in it. Faults must be quoted, never merely described. The Creator then rewrites to answer them. Two rounds, then the final story and whatever notes the Critic still has. A standing rule for the Gem: the Critic finds something every single round, the last one included. Run it now. The story: a girl who saves water at home. I teach this class in Kottayam and will read it aloud to them. When the final story is ready, export it to Google Docs so I can print copies for the class.
Then loop it
- The Critic went easy on the preaching rule - the closing paragraph is still a lecture. Critic: take only the ending and quote the preaching sentences. Creator: rewrite the ending alone.
- Now something your checklist cannot settle. Use Search to find what is published about English vocabulary levels expected in Class 3 under the Kerala state syllabus, link what you find, and tell me honestly if the answer is that no usable word list is published - in which case my own reading to the class is the only test that counts.
- Lovely. Add 5 simple comprehension questions and a Malayalam version of the story. Then save all of this as the Gem's instructions, including the one-weakness rule.
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.